lecture 17: logic modeling of cell signaling networks€¦ · biochemistry 49: 3216 [2010] j...

LOGIC MODELING OF

CELL SIGNALING NETWORKS

J Saez-Rodriguez – Molecular Systems Biology 5: 331 [2009]

MK Morris – Biochemistry 49: 3216 [2010]

J Saez-Rodriguez – Cancer Research 71: 5400 [2011]

MK Morris – PLoS Computational Biology 7: e1001099 [2011]

1

Central Topic:

extracellular ligand ‘cues’

‘signals’

behavior ‘response’

(phenotype)

Regulation of Mammalian Cell Behavior by Receptor-Mediated Signaling

[‘execution’– transcription / translation,

metabolism / synthesis, cytoskeleton / motors]

Source: Hanahan, Douglas, and Robert A. Weinberg. "The Hallmarks of Cancer." Cell 100, no. 1 (2000): 57-70.

2

Courtesy of Elsevier, Inc., http://www.sciencedirect.com. Used with permission.

http://dx.doi.org/10.1016/S0092-8674(00)81683-9

http://www.sciencedirect.com

Objective: Learn how cell signaling network cell / tissue

operation – in multi-pathway manner -- differs phenotypic behaviorbetween normal and disease state

or among various individuals

environmental context altered behavior of

cellular ��machines��

and ��circuits��

expression protein levels

DNA dynamicmRNA sequence protein operationsgene

variations / mutations

3

Example: Myriad -- and highly diverse -- genetic alterations (amplifications, deletions, mutations) across pancreatic tumors… (as well as in breast, colon, brain)

© American Association for the Advancement of Science. All rights reserved. This content is excludedfrom our Creative Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.Source: Jones, Siân, Xiaosong Zhang, et al. "Core Signaling Pathways in Human Pancreatic CancersRevealed by Global Genomic Analyses." Science 321, no. 5897 (2008): 1801-6.

[[JJoonneess eett aall..,, SScciieennccee ((22000088)) ]] !! 4

http://ocw.mit.edu/help/faq-fair-use/

http://dx.doi.org/10.1126/science.1164368


…but diverse mutations lead to dysregulation of a limited set of key pathways at protein level






Cell Signaling ��Circuitry��

Need to advance

from Metaphor

to Model

© Scientific American Library. All rights reserved. This content is excluded from our CreativeCommons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.Varmus, H., and R. A. Weinberg. "The Genetic Elements Governing Cancer: Tumor SuppressorGenes." Genes and The Biology of Cancer (1993): 101-9.

Varmus & Weinberg, Genes & the Biology of Cancer [1993] 7


Spectrum of Computational Modeling Methods

SPECIFIED ABSTRACTED

differential equations

Boolean/fuzzy logic, decision trees

Bayesian networks mutualmechanisms information

logic regression, clusteringinfluences

topology

relationships

��prior knowledge�� needed

8

Database Pathways Relevant No. Genes Format GeneGO 700+ 55 804 Table

PANTHER 165 14 1,025 SBML CellMap (NetPATH) 20 12 625 BioPAX / SIF

Reactome 1081 4 173 BioPAX / SIF NCI-PID 104 28 459 BioPAX / SIF KEGG 1000+ 8 564 -

SUMMARY 120 2,054

Pathway / Interactome Databases hold substantial prior knowledge

Pathway Databases (Nodes)

Interactome Databases (Edges) Database Type No. Edges Graph type i2D v1.71 Protein-Protein (Exp) 11,327 Undirected STRING Integrated Text mining 35,033 Mixture GeneGo Curated 11,994 Directed, Signed Cell Map Curated 12,933 Mixture NCI-PID Curated 14,58 Mixture

Reactome Curated 6,930 Mixture SUMMARY 68,067 Mixture

© Respective copyright holders. All rights reserved. This content is excluded from our CreativeCommons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.

9


 

 

Pathway / Interactome Databases hold substantial ��prior knowledge�� for integrative analysis of multi-pathway network effects; but, there is need to move forward from illustration to prediction

EREG HBEGF

ERBB4

© source unknown. All rights reserved. This content is excluded from our Creative

Shortcomings:

• Typically diverse withNode Source respect to specificity and≥ 2

GeneGo context – i.e.,KEGG NCI-PID cell type, genomic content,NetPATH and/or environmentalPANTHER Reactome conditions

• Do not readily permit ��input-output�� calculation of network operating behavior, and thus difficult to relate to

phenotype and/or interventions

Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.

© Respective copyright holders. All rights reserved. This content is excluded from our CreativeCommons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.

10



© source unknown. All rights reserved. This content is excluded from our CreativeCommons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.

Central Goal:

Establish methodology for converting from qualitative

cell pathway topology ��maps�� to quantitatively

computable network models

Approach:

Employ logic-based modeling framework, to train qualitative ��prior knowledge�� maps to

quantitative empirical data for system context and

multi-pathway comparisons of interest

11


More Detailed Insights from Stronger Modeling Analysis -- integrating empirical data with prior knowledge

using network logic approach

Network Logic ModelGeneric Pathway Map (e.g., Ingenuity) Boolean operators:

nodes (=compounds), AND / OR / NOTsigned directed edges (activation +, inhibition -)

AND A B C

E F

G

S

AND

OR

NOT

NOT

A B C

E

��!!

F

G��!!

+ ++

+ +

+

12

Primary Hepatocytes HepG2 Huh7 Hep3B FocusFocus


Example Study: Comparative Hepatocytic Cell SignalingNetwork Operation in Inflammation Context

hepatocellular lineshepatocellular lines

Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/. 13

[7 ligands + control]

X [7 inhibitors + control]

X 17 signals = ~ 1000

measurements for each

cell-type and time-point and

replicate


Multi-Pathway Phosphoproteomic Data – primary human hepatocytes, HepG2 hepatocellular line

LATE TRANSIENT

NO RESPONSE SUSTAINED

Time-points: 0, 30 min, 3 hrs


-- also cell death, proliferation index, and production of ~50 cytokines for each condition

14


Example Database Pathway Map

from literature curation, for network responses

to our cytokine and growth factor treatments

from Ingenuity supplemented by some literature

knowledge for key receptors

-- 82 nodes, 116 edges


15


Choose Submap

Process DataNormalize between 0,1Filter noise, saturation

STOP?

Compare experiment - simulation

Import Data

Define Pathway Mapfrom literature/Database

Perform Experiments

Analyzeresulting model

CellNetOptimizer

DataRail

YESNO

EGF TNF

Ikb

0

870

Evaluate model Sum deviations + Size

1

0

0

1

Dev

iatio

n

Import Map

AND OR

EGF TNF

EGF TNF

Ikb

Ikb

TNFR EGFR

PI3K

AKT

IKKab

Ikb

TNF EGF

TNF EGF

IKKab

Ikb

TNFR EGFR

PI3K

AKTIKKab

Ikb

TNF EGF

OR

TNFR EGFR

PI3K

AKT

IKKab

TNF EGF

AND

TNF EGF

IKKab

Ikb

OR

Ikb

Process MapCompress & Remove non-observables

X

Training Prior Pathway Map Knowledge on Context-Specific Empirical Signaling Data

Create Boolean Scaffold

XX


Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.


 

  

 

Automated Development of Logic Network Models from Fit of Generic Pathway Map to Experimental Data

as an Optimization Problem

Objective θ = θ f + α ⋅ θS

Size of modelFit to data

θ f = (Bikl M − Bikl

E )2 K =1

M

∑ l=1

S

∑ θS = ν k Pk k=1

n

∑ Relative

importance Fit vs. Size

Function

∈ [0,1)∈ {0,1}∈ {0,1}


! minimize Objective Function (θ) across model variants (P), ! trading off model-data error and model size;

! αα ascertained by Pareto optimum for false-positive vs false- negative trade-offs

! obtain family of best-fit models (within 1% of Objective Function optimum)

17


rsha10

Line

 

 

 

 

 

 

Automated Development of Logic Network Models from Fit of Generic Pathway Map to Experimental Data

as an Optimization Problem

Genetic Algorithm 1. Initialize a population of model variants (from Ingenuity

scaffold or from random scaffolds) 2. Evaluate objective function (model-vs-data error plus model-

size penalty) for each individual in the population 3. Generate next generation of population using Elite Survival,

Fitness Selection, Mutation, and Crossover 4. Assess whether stop criterion is fulfilled, or iterate back to

step 2 5. Model pruning to reduce model size without detriment to

model-vs-data error 6. 100 runs for each value of model-size penalty αα

18

Illustration for HepG2 cell line

– improvement in data fit from best-fit original

scaffold model to best-fit trained model

Training data fit to ~9% error, substantially improved from original scaffold model fit of

>45% error

19© source unknown. All rights reserved. This content is excluded from our CreativeCommons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.


Illustration for HepG2 cell line

– consensus model from fit of empirical data to initial prior knowledge

scaffold -- additional arcs needed

to improve model fit, support in literature though not in prior knowledge scaffold

-- arcs present in other cell line models but not in

HepG2

Courtesy of EMBO and Nature Publishing Group. License: CC-BY-NC-SA.Source: Saez‐Rodriguez, Julio, Leonidas G. Alexopoulos, et al. "Discrete Logic Modelling as a Means toLink Protein Signalling Networks with Functional Analysis of Mammalian Signal Transduction."Molecular Systems Biology 5, no. 1 (2009).

20

http://dx.doi.org/10.1038/msb.2009.87

http://dx.doi.org/10.1038/msb.2009.87

Model size is fairly insensitive to size penalty

Objective function = Fit of data (MSE) + α Size

200 0.2

160

0 10 10 10 10 Size penalty !

selected model-size penalty, for maximal predictive capability


Substantial number of scaffold arcs not supported

by hepatocyte data MS

E !

f

Obje

ctive function !

Siz

e !

S 0.15 120

80

40

0.1

0.05

00 -3-7 -5 -1

21


0% identifiability

* *

Models can only be partially identified -- thus model families are best outcome

Frequency of Arc Distribution for Error Tolerance-Related (i.e., beyond exptl uncertainty) Model Families

arc frequency


22


 

 

 

Trade-off between False Negatives and False Positives

• Receiver Operating Characteristic (ROC) curve [ratio of true positives (1-false negatives) vs. false positives] for different values of the size penalty α

• Optimal choice of size penalty (α=10-5) corresponds to most predictive model

• Extended model (i.e., with added arcs) decreases false negatives but increases false positives


23


24

Model Validation -- successful a priori

predictions of new test data

!  Used trained model to a priori predict effects of ligand combinations,

additional inhibitors, and inhibitor combinations

!  New test data predicted to within ~11% error,

comparable to ~9% for original training data

!  Can identify loci needing more detailed inquiry

0 0.5 1

Agree Disagree



Extension to comparison among hepatocellular lines -- phosphoproteomic data

© American Association for Cancer Research. All rights reserved. This content is excluded from ourCreative Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.Source: Saez-Rodriguez, Julio, Leonidas G. Alexopoulos, et al. "Comparing Signaling NetworksBetween Normal and Transformed Hepatocytes Using Discrete Logical Models."Cancer Research 71, no. 16 (2011): 5400-11. 25


http://dx.doi.org/10.1158/0008-5472.CAN-10-4453


Demonstration of benefit of cell type-specific models

Demonstration of capability to identify particular points

inviting further study

© American Association for Cancer Research. All rights reserved. This content is excluded from ourCreative Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.Source: Saez-Rodriguez, Julio, Leonidas G. Alexopoulos, et al. "Comparing Signaling NetworksBetween Normal and Transformed Hepatocytes Using Discrete Logical Models."Cancer Research 71, no. 16 (2011): 5400-11.

26




 

 

 

 

 

 

Best-Fit Boolean Logic Model Families

for Primaries versus Lines

• Arc width corresponds to proportion of best-fit models bearing it • Black arcs – all models in both primaries and HCC lines • Blue arcs – most or all primary models • Red arcs – most or all HCC line models • Gray arcs deleted from original scaffold • Dashed arc added to account for especially recalcitrant data





 

 

 


for Primaries versus Lines

• ~90% of original scaffold interactions were found in at least one best-fit model across families for all cell types • but only <10% were found both in most primary cell models and cell line models • multiple pathways are identifiable as dysregulated from normal to tumor lines





 

 

Model permits novel insights concerning

drug actions

Dashed arc added to fit data generated in presence of IKK inhibitor TPCA1 – two potential explanations: • IKK activity suppresses STAT3 activity downstream of JAK2; or • TPCA1 has off-target effect on JAK2





Experimental validation of model prediction that putative IKK inhibitor TPCA1 hits JAK2 as an off-target substrate (whereas BMS-345541 does not)


…perhaps providing an explanation for why TPCA1 has been found to be more efficacious for airway inflammation treatment than other IKK inhibitors

30




 


-- comparison among HCC Lines

• cell-type specificity of network operation is thus explicitly characterized – not only contrasting primaries to tumor lines but also disparities between different tumor lines


31




Cell types can be quantitatively clustered with respect to common edges

-- reasonable similarity to transcriptomic result





 

Detailed Primary-vs-Lines Comparison

• 8 edges are strongly disparate between primary

hepatocytes and the HCC lines





 

 

Detailed Primary-vs-Lines Comparison – insights gained

• Whereas EGFR leads to ERK activation in all cell

types, HSP27 is significantly activated

downstream of ERK only in primaries

• In the lines, HSP27 was activated more mildly and via p38 instead of via ERK

(Literature: HCC tumor

progression is associated with

decreased HSP27 activation --

despite HSP27 over-expression)





 


© American Association for Cancer Research. All rights reserved. This content is excluded from our

• In primaries Ikb phosphorylation

requires TNFa-NIK and activation of PI3K-JNK

(via TGFa or Ins), whereas in lines only TNFa-NIK is required

(Literature: HCC tumor

progression is associated with looser control

over NFkB-mediated

survival signals)

Creative Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.Source: Saez-Rodriguez, Julio, Leonidas G. Alexopoulos, et al. "Comparing Signaling NetworksBetween Normal and Transformed Hepatocytes Using Discrete Logical Models."Cancer Research 71, no. 16 (2011): 5400-11. 35




 


© American Association for Cancer Research. All rights reserved. This content is excluded from our

• GSK3 phosphorylation by

Akt (leading to nuclear activation of pro-mitotic factors)

is induced by Insulin in lines but not in

primaries

(Literature: IRS1 is over-expressed in

HCC, potentially shifting Insulin-

induced signaling from IRS2-mediated metabolism to proliferation)

Creative Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.Source: Saez-Rodriguez, Julio, Leonidas G. Alexopoulos, et al. "Comparing Signaling NetworksBetween Normal and Transformed Hepatocytes Using Discrete Logical Models."Cancer Research 71, no. 16 (2011): 5400-11. 36




These same three pathways have been implicated in

combination kinase therapy for HCC

IKK Akt p38

© American Association for Cancer Research. All rights reserved. This content is excluded fromour Creative Commons license. For more information, see http://ocw.mit.edu/help/faq-fair-use/.Source: Pritchard, Justin R., Benjamin D. Cosgrove, et al. "Three-kinase Inhibitor CombinationRecreates Multipathway Effects of a Geldanamycin Analogue on Hepatocellular Carcinoma CellDeath." Molecular Cancer Therapeutics 8, no. 8 (2009): 2183-92.

37


http://dx.doi.org/10.1158/1535-7163.MCT-08-1203



��Constrained Fuzzy Logic�� Framework

allows Analog Model rather than Digital

Courtesy of Morris et al. License: CC-BY.Source: Morris, Melody K., Julio Saez-Rodriguez, et al. "Training Signaling Pathway Maps to Biochemical Datawith Constrained Fuzzy Logic: Quantitative Analysis of Liver Cell Responses to Inflammatory Stimuli."PLoS Computational Biology 7, no. 3 (2011): e1001099. 38

http://dx.doi.org/10.1371/journal.pcbi.1001099


HepG2 Constrained Fuzzy Logic Network Model (again consensus family)

Extracellular cues

Protein signals

Courtesy of Morris et al. License: CC-BY.Source: Morris, Melody K., Julio Saez-Rodriguez, et al. "Training Signaling Pathway Maps to Biochemical Datawith Constrained Fuzzy Logic: Quantitative Analysis of Liver Cell Responses to Inflammatory Stimuli."PLoS Computational Biology 7, no. 3 (2011): e1001099.

IGF1

Ras PI3K

IL1a

TRAF6

IL6

STAT3

LPS TGF!TNF!

MKK4

MAP3K7

JNK1/2

MEK1/2MAP3K1 Akt

IKK

p53 I"B

p38

0.18 ± 0.08

0.38 ±0.43 ±

p70s6

p90RSKMSK1/2

IRS1sHsp27

mTOR

Gsk3 CREBHistH3c-Jun

0.6 ± 0.15

0.41 ± 0.04 0.06 ± 0.01 0.2 ± 0.08

0.11 ± 0.05

0.59 ±0.17

0.450.15*

0.13 ± 0.08

0.38 ± 0.09 0.32 ± 0.13

0.2 ± 0.08

0.5 ± 0

0.68 ± 0.18

0.3 ±0.07

0.49 ± 0.15

0.7

5 ±

0.1

8

0.49 ± 0.16

0.1

9 ±

0.0

5

0.57 ± 0.19

0.37 ±0.13

0.17

0.66 ±0.14

0.6

± 0.1

3

0.8 ± 0.16

0.20.34 ± 0.18

0.5

± 0.1

7

0.28 ± 0.11

0.38 ± 0.16

0.62 ±

0.41 ± 0.17 0.68 ±

0.11

*

Intensity of arc = likelihood of connection

Numerical descriptor = upstream-downstream effect strength

* = new arcs not identified by Boolean model 39



40

0.2

0.6 ± 0.15

0.18 ± 0.08

0.59 ± 0.17

0.2 ± 0.08

0.3 ± 0.07

0.7

5 ±

0.1

8

0.1

9 ±

0.0

5

0.57 ± 0.19

0.37 ± 0.13

0.38 ± 0.17

0.66 ± 0.14 0.45

0.6

± 0

.13

0.5

± 0.1

7

0.62 ± 0.2

0.68 ± 0.15

0.43 ± 0.11

1.2

1

0.8

0.6

0.4

Example Results for Quantitative Cell Circuit Logic -- downstream ��child�� node versus upstream ��parent�� nodes

Red points: experimental values Gray points: averaged-model predictions Gold points: individual model predictions

TNF! IL1a LPS TGF! IGF1 0.38 ± 0.09

0.32 ± 0.13

0.13 ± 0.08 0.41 ± 0.04 0.06 ± 0.01

IL6

0.5 ± 0

0.2 ± 0.08 PI3KRasTRAF6

0.41 ± 0.17

0.11 ± 0.05

MAP3K7 MAP3K1 Akt MEK1/2

MKK4 IKK 0

-0.2 JNK1/2 p38 1

0.49 ± 0.150.9 0.8 1 0.8 ± 0.16

0.7 0.9 MSK1/2 p90RSK

0.34 ± 0.18

mTOR

0.68 ± 0.18 0.49 ± 0.16

0.6 0.8 0.70.5 0.6 0.38 ± 0.16

0.4 0.50.3 0.4 0.28 ± 0.11MEK1/2 0.2 0.3 c-Jun Hsp27 p53 I"B Gsk3 HistH3 CREB IRS1s p70s6 STAT30.20.1

0 0 p380.1

Courtesy of Morris et al. License: CC-BY.Source: Morris, Melody K., Julio Saez-Rodriguez, et al. "Training SignalingPathway Maps to Biochemical Data with Constrained Fuzzy Logic:Quantitative Analysis of Liver Cell Responses toInflammatory Stimuli."PLoS Computational Biology 7, no. 3 (2011): e1001099.

CREB

New test data fell within one standard deviation

of predictions across all conditions

© source unknown. All rights reserved. This content is excludedfrom our Creative Comm ons license. For mor e information, seehttp://ocw.mit.edu/help/faq-fair-use/.





Model family precision generally presages accuracy

Courtesy of Morris et al. License: CC-BY.Source: Morris, Melody K., Julio Saez-Rodriguez, et al. "Training Signaling Pathway Maps to Biochemical Datawith Constrained Fuzzy Logic: Quantitative Analysis of Liver Cell Responses to Inflammatory Stimuli."PLoS Computational Biology 7, no. 3 (2011): e1001099. 41



MIT OpenCourseWarehttp://ocw.mit.edu

7.91J / 20.490J / 20.390J / 7.36J / 6.802J / 6.874J / HST.506J Foundations of Computational and Systems BiologySpring 2014

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

http://ocw.mit.edu

http://ocw.mit.edu/terms

lecture 17: logic modeling of cell signaling networks€¦ · biochemistry 49: 3216 [2010] j...

Documents