Objective Bayesian Nets for Integrating Cancer Knowledge

1
Objective Bayesian Nets for Integrating Cancer
Knowledge
Sylvia Nagl PhD Cancer Systems Biology
Biomedical Informatics UCL London
2
caOBNET Overview

• Knowledge integration by objective Bayesian
  networks (obNETS)
• Maximum entropy method
• An integrated clinico-genomic obNET for breast
  cancer
• Conclusions

3
Bayesian networks

• Graphical models
• directed and acyclic graph (DAG)
• Joint multivariate probability distribution
• with conditional independencies between
  variables
• Given the data, optimal network topology can be
  estimated
• heuristic search algorithms and scoring
  criteria
• Statistical significance of edge strengths
• Bayesian methods
• bootstrapping

Apolipoprotein E gene SNPs and plasma apoE level
Rodin Boerwinkle 2005
4
Knowledge integration

• Cancer treatment decisions should be based on all
  available knowledge
• Knowledge is complex and varied
• Patient's symptoms, expert knowledge, clinical
  databases relating to past patients, molecular
  databases, scientific papers, medical informatics
  systems
• Generated by independent studies with
• diverse protocols

5
Knowledge integration

• Diverse data types
• Genomic, transcriptomic, proteomic, SNPs, tissue
  microarray, histopathology, clinical etc.
• New data types, e.g., epigenetic data
• All data types capture different characteristics
  of a dynamic complex system
• At different spatial and temporal scales
• Cell, tumour, patient, and therapeutic system of
  patient-therapy interactions
• How can this disparate data be used for an
  integrated understanding on which to base our
  actions?

6
Objective Bayesianism

• Data and knowledge impinge on belief we try to
  find a coherent set of beliefs with best fit
• Beliefs based on undefeated items of knowledge
• In case of conflict, try to find compromise
  beliefs
• Objective Bayesianism offers a formalism for
  determining the beliefs that best fit background
  knowledge
• Applying Bayesian theory, an agents degree of
  belief should be representable by a probability
  function p
• Empirical knowledge imposes quantitative
  constraints on p
• Represented in an obNET (learnt from database)

7
obNETS for prediction

• Standard algorithms can be used to calculate the
  probability of a specific outcome
• A direct link between variables may suggest a
  causal connection

8
Bayesian networks

• Can BNs be integrated?
• Spanning genetic/molecular and clinical levels
• obNETS offer a principled path to knowledge
  integration

9
Maximum entropy principle

• Adopt p, from all those that satisfy the
  constraints, that are maximally equivocal
• Williamson, J.(2002) Maximising Entropy
  Efficiently.
• Williamson, J. (2005a) Bayesian Nets and
  Causality.
• Williamson, J. (2005b) Objective Bayesian nets.
• www.kent.ac.uk/secl/philosophy/jw/

10
Example

• Two items of empirical knowledge may conflict
• Study 1 Cancer will recur in 50 of patients
  with given set of characteristics
• Degree of belief in recurrence in individual
  patient 0.5
• Study 2 Frequency of recurrence is 30
• Degree of belief will be constrained to closed
  interval 0.3,0.5
• In general
• Belief function will lie within a closed set of
  probability functions
• There will be a unique function that maximises
  entropy

11
obNet integration
12
obNet integration
Original obNETs provide probability distributions
13
obNET integration
14
obNET integration
15
obNET integration

• n number of nets

16
obNET integration
Maximum entropy principle If CPTs for merged
nodes disagree on probabilities, assign closed
interval and take least committal value in that
range
17
obNET integration Proof of principle

• Two obNETs from breast cancer knowledge domain
• Genomic Comparative genome hybridisation (CGH)
  data - progenetix database
• Subset of bands with 3 or more genes implicated
  in tumour progression and response to cytotoxic
  therapies (28 bands)
• Clinical American Surveillance, Epidemiology and
  End results (SEER) database

18
Clinical and genomic nets (Hugin 6.6)
SEER database 4731 cases progenetix
database 28 bands/502 cases
?
19
obNet integration
obNet learnt from 2nd progenetix dataset - 119
cases with clinical annotation (lymph node
status, tumour size, grade)
CPT
22q12 -1 0 1 LN0 0.148 0.5 0.148
1 0.852 0.5 0.852
20
Additional empirical knowledge
chr. 22
Fridlyand et al. 2006
21
obNet integration
chr. 22
Fridlyand et al. 2006
CPT
22
obNet integration
chr. 22
Fridlyand et al. 2006
CPT
23
Metastasis-associated genes
KREMEN1 MYH9
cadherin11
CD97
BMP7, ELMO2, BCAS1, BCAS4, ZNF217
24
KREMEN1
Howard et al., 2003
Biological knowledge suggests possible causal
link (in context of whole obNET HR status!)
25
Knowledge integration
Multi-scale obNETs
Cancer clinical data epidemiology
Translation of clinical data to genomics research
Predictive markers
Molecular profiling of tumours
26
Acknowledgements

• Jon Williamson (Philosophy, Unversity of Kent)
• www.kent.ac.uk/secl/philosophy/jw/
• Matt Williams (Cancer Research UK)
• Nadjet El-Mehidi (Cancer Systems Biology, UCL)
• Vivek Patkar (Cancer Research UK)
• Contact s.nagl_at_ucl.ac.uk

27
obNET integration Proof of principle

• Two obNETs
• Non-independent rearrangements at chromosomal
  locations in breast cancer from comparative
  genome hybridisation (CGH) data - progenetix
  database
• Subset of bands with 3 or more genes implicated
  in tumour progression and response to cytotoxic
  therapies (28 bands)
• Probabilistic dependencies between clinical
  parameters from the American Surveillance,
  Epidemiology and End results (SEER) database

28
HR status link
29
Genomic systems

• Genomes are dynamic molecular systems
• Selection acts on unstable cancer genomes as
  integrated wholes, not just on individual
  oncogenes or tumour suppressors.
• A multitude of ways to solve the problems of
  achieving a survival advantage in cancer cells
• Irreversible evolutionary processes
• Randomness of mutation
• Modularity and redundancy of complex systems

30
Genome-wide rearrangements

• Can we identify probabilistic dependency networks
  in large sample sets of genomic data from
  individual tumours?
• If so, under which conditions may these be
  interpreted as causal networks?
• Can we identify probabilistic dependency networks
  involving molecular and clinical levels?

31
Systems Biology and Causation

• Profound conceptual challenge regarding physical
  causation in complex biological systems
• Mutual dependence of physical causes
• The biological relevance of any factor, and
  therefore the information it conveys, is
  jointly determined, frequently in a statistically
  interactive fashion, by that factor and the
  system state (Susan Oyama, The Ontogeny of
  Information, 2000)
• The influence of a gene, or a genetic mutation,
  depends on the context, such as availability of
  other molecular agents and the state of the
  biological system, including the rest of the
  genome

32
System state
agents
Cell networks are dynamically instantiated
genes for components are switched on or off in
response to signals and cell state
33
System state
Cell networks are reconfigured in response to
changes in environment or cells internal state
34
System state
Cell computation networks are reconfigured in
response to changes in environment or cells
internal state
35
Cancer Genome instability re-programs cell
networks
Selection for increased proliferation,
resistance, invasiveness etc. Driven by tumour
cell tissue interactions
36
Genome-wide rearrangements

• Can we identify probabilistic dependency networks
  in large sample sets of genomic data from
  individual tumours?
• Can we identify probabilistic dependency networks
  involving molecular and clinical levels?

37
Proof of principle

• Screen the whole genome for chromosomal
  abnormalities in one experiment
• Cytogenetics
• Comparative genomic hybridization (CGH)
• Fluorescence in situ hybridization (FISH) and
  multicolour fluorescence in situ hybridization
  (MFISH)
• Detection of allelic instabilities, loss of
  heterozygosity (LOH)