Cerebral: Visualizing Multiple Experimental Conditions on a Graph with Biological Context

1
Cerebral Visualizing MultipleExperimental
Conditions on a Graph with Biological Context

• M.Sc Thesis Presentation
• Aaron Barsky
• Supervisor Tamara Munzner
• May 21st, 2008

2
Cerebral

• Collaboration with systems biologists
• Hancock innate immunity laboratory
• Integrate
• System model (Graph)
• Experimental measurements

3
Outline

• Cellular systems biology
• Design decisions
• Related work
• Constrained simulated annealing graph layout
• Interactive data exploration video
• Conclusions / Future work

4
Biomolecular interactions are selective

• Cell densely packed with biomolecules
• Interactions rare
• Model interactions as a graph

Image from Nature Publishing group
5
Systems biology model

• Graph G V, E
• V proteins, genes, DNA, RNA, tRNA, etc.
• E interacting molecules
• Reactions, information exchange, controls

6
Graph summarizes extensive lab work

• Graphs extracted from database
• Each edge summarizes experimental evidence
• TIRAP an adapter molecule in the Toll signaling
  pathway. Horng T, Barton GM, Medzhitov R.
• Mal (MyD88-adapter-like) is required for
  Toll-like receptor-4 signal transduction. Fitzgera
  ld KA, Palsson-McDermott EM, Bowie AG, Jefferies
  CA, Mansell AS, Brady G, Brint E, Dunne A, Gray
  P, Harte MT, McMurray D, Smith DE, Sims JE, Bird
  TA, O'Neill LA.

7
Models are dynamic

• No official summary source
• Changes with each publication
• Exploration of diagrams useful
• Choose scope to manage complexity
• Reactions associated with a biomolecule
• Reactions associated with a system
• Reactions associated with an organism

8
TLR4 context E74, V54
9
Immune system context E1263, V760
10
Immune system context E1263, V760
11
Human cell (E50,000, V10,000)
12
Model interprets experiments.Experiments refine
model.

• Systems biologists
• Conduct experiments on cells
• Interpret results in current model
• Propose modifications to the model

13
Microarray experiments measure gene expression
level

• Cells express genes to create proteins
• Proteins are specialized tools
• Associate real value with each graph node

Image www.immb.forth.gr
14
LL-37 Example

• Suspect LL-37 helps reduce inflammation
• Drug not part of model
• Conduct experiment
• Treat some cells with LL-37
• Controls untreated
• Expose all cells to bacteria
• Measure gene expression over 4 time points

15
Experiment results

• Context TLR4 immune response

16
Cerebral
17
Thesis contributions

• Cerebral Interactive exploration tool
• Views multiple experimental conditions
• In context of graph model
• Facilitates comparison between pairs of
  conditions
• Graph layout algorithm
• Uses biological meta-data

18
Outline

• Cellular systems biology
• Design decisions
• Related work
• Constrained simulated annealing graph layout
• Interactive data exploration video
• Conclusions / Future work

19
Many visualization options

• Input
• Graph G V, E
• Descriptive meta-data for each v in V
• Labels, biological attributes
• Sets of experimental results
• Multiple float values associated with each v in V
• Output
• Graphical representation

20
Our choices

• Graph layout guided by biological meta-data
• Small multiple views for experimental conditions
• Parallel coordinates for a measurement- driven
  view

21
Traditional graph layout

• Given graph GV,E
• Create layout in 2D plane

Circular (Six and Tollis, 1999)
Force-directed (Fruchterman and Reingold, 1991)
Hierarchical (Sugiyama 1989)
22
Good layout criteria

• Short edges
• Minimal edge crossings
• Minimal node-edge overlap
• Compactness
• Symmetry
• Empirical Evaluation of Aesthetics-based Graph
  Layout (Purchase, 2002)
• Many criteria NP hard

23
Biologists found existing layouts unsuitable

• Generic layout criteria form unexpected groupings
• Thats weird. Why is that transcription factor
  beside that cell surface protein?
• Biologists want graph layout to encode
  biological structure

24
Biological cells divided by membranes
Image courtesy of Dr.G Weaver

• Interactions generally occur within a
  compartment
• Crossing membranes interesting

25
Hand-drawn diagrams

• Cellular location encoded spatially

26
Cerebral spatial encoding

• Similar to hand-drawn
• Spatial position reveals
• Location in cell
• Function

27
Small multiple views for experimental conditions

• One graph instance per condition
• Each graph coloured according to the condition
• Tufte, 1990

28
Animation over time
29
Visual memory poor

• Matthew Plumlee and Colin Ware. Zooming versus
  multiple window interfaces Cognitive costs of
  visual comparisons. Proceedings of the ACM SIGCHI
  Conference on Human Factors in Computing
  Systems,13(2)179-209, 2006.

• Barbara Tversky, Julie Bauer Morrison, and
  Mireille Betrancourt. Animation can it
  facilitate? International Journal of
  Human-Computer Studies, 57(4)247-262, 2002.

30
Embedded glyphs

• Embed multiple conditions as a chart in the node
• Good detail in local view
• Westenberg 2008

31
Glyphs invisible in global view

• Westenberg, 2008

32
Saraiya study

• Purvi Saraiya, Peter Lee, and Chris North.
  Visualization of graphs with associated
  timeseries data, 2005.
• Compared 4 interfaces for analyzing expression
  data in a graph
• Animated coloured nodes outperformed glyphs
• Multiple linked views improve accuracy
• Aim to do better by distributing over space vs.
  over time

33
Parallel coordinates for ameasurement driven view

• Each experimental condition is an axis
• Each node in the graph is a line

34
Clusters indicate similar function?

• Data driven hypothesis
• Same pattern of gene expression same role in
  cell
• Parallel coordinates alone untrustworthy
• Data noisy
• Different clustering algorithm different
  results

35
Linked graph and clustering aid exploration
36
Outline

• Cellular systems biology
• Design decisions
• Related work
• Constrained simulated annealing graph layout
• Interactive data exploration video
• Conclusions / Future work

37
Related Work

• Systems biology graph viewers
• Constrained graph layout

38
Systems biology graph visualization systems

• Cytoscape (Shannon et al. 2003)
• VisANT (Hu et al. 2004)
• GeneSpring (Silicon Genetics)
• GenMapp (Dalquist et al. 2002)
• Graph layout without biological context
• Overlay only a single condition at a time

VisANT (Hu et al. 2004)
39
Multiple coordinated views

• Multiple linked views of experiment data
• HCE (Seo and Schneiderman 2002)
• SpotFire (Tibco SpotFire)
• No graph view

40
Constrained graph drawing

• Force directed
• Contain with repulsive walls
• Graph drawing by force directed placement
  (Fruchterman and Reingold, 1991)
• A constrained, force-directed layout algorithm
  for biological pathways (Genc and Dogrusoz 2004)
• Force balancing a challenge
• Parameter tweaking
• Brittle

41
Quadratic programming

• Numerical approaches
• Constrained graph layout (He, Marriott 1998)
• IPSep-CoLa An incremental procedure for
  separation constraint layout of graphs (Dwyer,
  Marriott 2006)
• Handles separation constraints
• Ideal edge length parameter
• Requires tweaking
• Does not optimize edge crossings

42
Graph layout with simulated annealing

• Simple and flexible
• Drawing graphs nicely using simulated annealing
  (Davidson and Harel 1996)
• Automatic drawing of biological networks using
  cross cost and subcomponent data (Kato and
  Nagasaki 2005)
• Historically slow O(V3)
• Our system, Cerebral expected O(EvV)

43
Outline

• Cellular systems biology
• Design decisions
• Related work
• Constrained simulated annealing graph layout
• Interactive data exploration video
• Conclusions / Future work

44
Graph layout constraints

• Restrict node placement to band according to
  subcellular localization
• Cluster activated proteins by function
• Optimize edge length, crossings, etc..

45
Simulated annealing search

• Choose a random graph layout
• Repeat until cool
• Repeat O(N)
• Move a random node to a new position
• Score new position with evaluation function
• If improved
• accept change
• Else
• accept change with probability 1

Reduce temperature
46
Adapting SA to constraints

• Hard constraints
• Layer by subcellular localization
• Soft constraints
• Minimize
• Edge length
• Edge-edge crossings
• Node-edge crossings
• Distance to biologically similar neighbours

Extracellular
Plasma membrane
Cytoplasm
Nucleus
47
Soft constraint violation evaluation is frequent

• Must be efficient
• Innermost loop
• 50 cooling cycles 30N nodes
• 1500N evaluations

48
Discretization key to efficiency

• Limit node positions to grid centers
• Uniform grid (Akman et al.,1989)

49
Clean, regular layouts

• Room for labels

50
No node overlaps

• Overlaps in dense areas of force directed
  algorithms

• Cerebral Impossible by construction
• - No cost to evaluate

51
Calculations with L1 distance

• Manhattan or L1 distance
• Cheap, integer only arithmetic
• Measures
• Edge length
• Distance to functional neighbours

52
High speed edge crossing count estimation

• Edge crossing count could be very expensive
• Count everything O(E2) or O(E log E)
• Count just the moved node O(deg(n)E)
• Over 98 of unoptimized time
• Cerebral Good estimate
• O (vV)
• Integer only

53
Quickly find cells with modified Bresenhams
algorithm

• Modified Bresenhams
• Green classic
• Purple additional corners

• Update grid cell each time edge is moved

54
Track edges in each grid cell

• Cell stores edge count

• Add up edges in cells a line passes through
• No expensive line intersection tests
• Upper bound estimate

55
High angular resolutionaids reading of edge
crossings

• Exact intersection test 3 crossings

• Approx. intersection test 12 crossings

56
Cerebral TLR4 (E74, V57) Time2.9 sec
57
Force-directed TLR4 Time lt1 sec
58
IPSep-CoLa Time1.3 sec
59
Cerebral innate immunity (V1263, N760) Time62
sec
60
Force-directed Time 64 sec
61
IPSep CoLa Time 296 sec
62
Cerebral graph layout

• Shows subcellular localization through spatial
  positioning
• Groups response proteins by biological function
• Requires no user specified parameters
• Runs in expected time O(EvV)
• A few minutes for 1000 nodes and edges

63
Outline

• Cellular systems biology
• Design decisions
• Related work
• Constrained simulated annealing graph layout
• Interactive data exploration
• Conclusions / Future work

64
Usable, but complex
65
Video
66
Outline

• Cellular systems biology
• Design decisions
• Related work
• Constrained simulated annealing graph layout
• Interactive data exploration video
• Conclusions/ Future work

67
Released in two stages

• Cerebral 1.0 (2007)
• Biologically based graph drawing only
• Announced in a Bioinformatics Application Note
• Cerebral 2.0 (now)
• Multiple experiment viewing with small multiple
  and parallel coordinate views

68
Implemented as a plugin for Cytoscape

• Cytoscape (Shannon et al. 2003)
• Open source systems biology tool
• Provides model/attribute management
• Replaced standard graph renderer
• Biologically based graph layout
• Added small multiple and parallel coordinate views

69
Biologists are using Cerebral

• Published Cerebral-created diagrams to
  communicate results
• Matthew D Dyer, T. M Murali, and Bruno W Sobral.
  The landscape of human proteins interacting with
  viruses and other pathogens. PLoS Pathogens,
  4(2)e32, 2008.

70

• Liqun He et al. The glomerular transcriptome and
  a predicted protein-protein interaction network.
  Journal of the American Society of Nephrology,
  19(2)260-268, 2008.

71
InnateDB links to Cerebral

• Integrated as a visualization system for InnateDB

72
Future Work
73
Visual scaling

• Human cell (V10,000, E 50,000) Time 6199 sec

74
Support Organelles

• Membrane-bound regions within Cytoplasm
• Mitochondria
• Lysosomes
• Vesicles

Mitochondria
Vesicles
Cytoplasm
75
Biology-specific clustering

• Replace k-means with biology specific clustering
  algorithm

76
Indicate data-mining bias

• Exploration without hypothesis
• Is pattern a signal?
• Measure significance of pattern given
• Graph size
• Connectivity of members
• Number of experiments

77
Conclusion

• Cerebral
• Visualizes experimental data from multiple
  conditions simultaneously
• Allow interactive exploration of the data
• Uses biological meta-data to guide the graph
  layout

78
Acknowledgements

• Funding
• Agilent Technologies
• Robert Kincaid
• Hancock lab members
• Jennifer Gardy, David Lynn, Bob Hancock
• Supervisor
• Tamara Munzner
• Information visualization group
• - Stephen Ingram, Peter McLachlan, Dan
  Archambault, Heidi Lam, James Slack, and Ciaran
  Llachlan Leavitt

79
Yeast cell cycle data

• Measure gene expression levels at 24 time points
  in yeast
• Through a cell divide cycle
• Observe

80
(No Transcript)
81
(No Transcript)
82
(No Transcript)
83
(No Transcript)
84
Principles of perception

• Group objects by colour/shape/size
• Group by rows, not columns
• Automatically by visual system
• Information Visualization - Perceptions for
  Design. Ware (2004)

85
Spatial position overrules colour/shape/size
groupings

• Automatically view 3 groups of mixed objects
• With effort can group by
• Size
• Shape
• colour

86
Connectedness even stronger than position

• Each group would be perceived differently without
  the connecting line

• Information Visualization - Perceptions for
  Design. Ware (2004)

87
(No Transcript)