Title: Towards a general computational method for robustness analysis Grgory Batt with Aurlien Rizk, Franoi
1Towards a general computational method for
robustness analysis Grégory Batt with
Aurélien Rizk, François Fages, Sylvain
SolimanContraintes project team, INRIA
Paris-Rocquencourt
April 14, 2009
2Robust behaviour of transcriptional cascade
synthetic transcriptional cascade
Hooshangi et al, PNAS, 05
- Design problem obtain a rapid and high-amplitude
response to aTc
3Robust behaviour of transcriptional cascade
synthetic transcriptional cascade
Hooshangi et al, PNAS, 05
- Design problem obtain a rapid and high-amplitude
response to aTc
Could be used as a timer enforce correct
sequencing of events
- But
- cells subject to environmental fluctuations
- robustness capacity of a system to maintain a
function in the face of perturbations - ? explore large ranges of parameter values and
determine robustness of cell response
4Describing the functioning of biological systems
- Robustness is threefold definition system,
function, perturbation - Large variety of function of interest ad hoc
definitions - growth signatures
- steady state values
- expression patterns
- presence, frequence, amplitude and phase of
oscillations - Definition of robustness as mean functionality
- where is probability of perturbation p and
is "evaluation" function - Robustness computation/estimation difficult to
automate - General problem with no general formalism
5Temporal logic as specification language
- Temporal logics are general-purpose languages for
specifying dynamical properties of discrete
transition systems - Paves the way for developing general methods for
robustness estimate - Linear time logic (LTL) syntax
- finite set of atomic propositions
- usual logical operators
- temporal operators
- Semantics of LTL formulas defined over traces
-
- Pnueli, FOCS 77
6Biological properties formalized in LTL
- Formalizing properties of (finite) numerical
traces in LTL - But true/false valuation of TL not adapted to
robustness analysis
7Biological properties formalized in LTL
- Formalizing properties of (finite) numerical
traces in LTL - But true/false valuation of TL not adapted to
robustness analysis
8Biological properties formalized in LTL
- Formalizing properties of (finite) numerical
traces in LTL - But true/false valuation of TL not adapted to
robustness analysis - Need to quantify how far is the system from
verifying the specification
9Validity domain of QFLTL formulae
- Validity domainD?(T) set of values of
variables in formula ? making it true on trace
T Fages and Rizk, CMSB'07 -
10Continuous valuation of LTL formulae
- Violation degree vd(T,?) distance between
specification and validity domain
Rizk et al, CMSB'08
11Robustness definition using violation degree
- System's evaluation function defined using
satisfaction degree - Instantiation of Kitano's definition of
robustness - Temporal-logic based definition generic approach
and computational method - Computational estimation of robustness
12Various notions of robustness
- Robustness average functionality
-
-
same mean behaviour
but significant difference between mean and
nominal behaviour
Ex von Dassow et al.
13Various notions of robustness
- Robustness average functionality
-
- Relative robustness average and nominal
functionality ratio -
same mean behaviour
but significant difference between mean and
nominal behaviour
14Various notions of robustness
- Robustness average functionality
-
- Relative robustness average and nominal
functionality ratio - same relative robustness different
robustness
same mean behaviour
but significant difference between mean and
nominal behaviour
Ex Gonze et al.
15Various notions of robustness
- Robustness average functionality
-
- Relative robustness average and nominal
functionality ratio - same relative robustness different robustness
- Formal definitions discriminate between different
robustness notions
same mean behaviour
but significant difference between mean and
nominal behaviour
16Computation of validity domain
?
17Computation of validity domain
- Computation by induction on the trace and on the
formula -
?
18Computation of validity domain
- Computation by induction on the trace and on the
formula -
?
19Computation of validity domain
- Computation by induction on the trace and on the
formula - D?(T) computation unions and intersections of
polytopes/orthotopes -
?
20Implementation in BIOCHAM
- BIOCHAM is a modeling environment for analysis of
biochemical systems - New features using satisfaction degree of
temporal logic formula - robustness computation wrt versatile
specifications? promotes genericity - parameter search wrt high level specifications
(currently up to 50 parameters)? ? avoids over
specification
http//contraintes.inria.fr/BIOCHAMRizk et al,
CMSB'08
21Application design of transcriptional cascade
synthetic transcriptional cascade
Hooshangi et al, PNAS, 05
- Find parameters for which a fast and
high-amplitude response to aTc is robustly
obtained
- Approach
- specify expected behavior as TL formula
- develop ODE model
- find a perturbation model
-
- ? explore large ranges of parameter values
and compute robustness of cell response
22Describing timed behavior of cascade
1) Specification fast and high-amplitude
response to aTc
23Describing timed behavior of cascade
1) Specification fast and high-amplitude
response to aTc
"well timed" if t1gt150, t2lt450 and ?tlt150
24Describing timed behavior of cascade
1) Specification fast and high-amplitude
response to aTc
"well timed" if t1gt150, t2lt450 and ?tlt150
2) ODE model with Hill functions
25Selection of perturbation model
- 3) Perturbation model
- Many kinds of perturbations are considered SDEs
with additive/multiplicative noise, random ODEs
with (log-)normally distributed parameters...
26Selection of perturbation model
- 3) Perturbation model
- Many kinds of perturbations are considered SDEs
with additive/multiplicative noise, random ODEs
with (log-)normally distributed parameters... - Which perturbation model is most faithful to
reality? - Compare predicted and observed coefficients of
variations for various fluorescence intensities
27Selection of perturbation model
- 3) Perturbation model
- Many kinds of perturbations are considered SDEs
with additive/multiplicative noise, random ODEs
with (log-)normally distributed parameters... - Which perturbation model is most faithful to
reality? - Compare predicted and observed coefficients of
variations for various fluorescence intensities - LogN parameter distribution provides
experimentally-consistent perturbation model
experimental measurement
coefficient of variations as function of mean
fluorescence
Hooshangi et al, PNAS, 05
computed using ODEs with LogN parameters
28Satisfaction degree and robustness landscapes
- 2D landscapes of satisfaction degree and
robustness (16D) - Large variations (2 orders of magnitude) of 2
parameters? - Parameter variations may alter correct
functioning of system - Robustness and relative robustness provide
important design information - identification of regions of robust behaviour
- identification of regions of discrepancy between
reference and mean behaviours - 8D robustness analysis
- Most robust parameters close to experimental best
fit parameters
robustness
satisfaction degree
relative robustness
29Robustness analysis by variance decomposition
- Determination of how each parameter globally
affects robust functioning using variance
decomposition methods
(first order) global sensitivity index
Saltelli et al, Wiley, 04
30Robustness analysis by variance decomposition
- Determination of how each parameter globally
affects robustfunctioning using variance
decomposition methods - Parameter contributions to robustness of temporal
behaviours
(first order) global sensitivity index
Saltelli et al, Wiley, 04
first and second order global sensitivity indices
31Robustness analysis by variance decomposition
- Determination of how each parameter globally
affects robustfunctioning using variance
decomposition methods - Parameter contributions to robustness of temporal
behaviours - parameters closely-related to output value
generally have higher sensitivities
(first order) global sensitivity index
Saltelli et al, Wiley, 04
32Robustness analysis by variance decomposition
- Determination of how each parameter globally
affects robustfunctioning using variance
decomposition methods - Parameter contributions to robustness of temporal
behaviours - parameters closely-related to output value
generally have higher sensitivities - sensitivity of expression rate of eyfp repressor
cI higher than that of basal expression rate of
eyfp
(first order) global sensitivity index
Saltelli et al, Wiley, 04
33Robustness analysis by variance decomposition
- Determination of how each parameter globally
affects robustfunctioning using variance
decomposition methods - Parameter contributions to robustness of temporal
behaviours - parameters closely-related to output value
generally have higher sensitivities - sensitivity of expression rate of eyfp repressor
cI higher than that of basal expression rate of
eyfp - relatively low influence of degradation parameter
? and input concentration uaTC
(first order) global sensitivity index
Saltelli et al, Wiley, 04
34Conclusion
- Instantiation of Kitano's definition of
robustness to temporal logic setting - ? general framework for robustness analysis
- ? implemented in publicly-available tool BIOCHAM
35Conclusion
- Instantiation of Kitano's definition of
robustness to temporal logicsetting - ? general framework for robustness analysis
- ? implemented in publicly-available tool BIOCHAM
- Temporal logic provides rich conceptual
environment formethodological developments
36Conclusion
- Instantiation of Kitano's definition of
robustness to temporal logicsetting - ? general framework for robustness analysis
- ? implemented in publicly-available tool BIOCHAM
- Temporal logic provides rich conceptual
environment formethodological developments - Application to design of synthetic
transcriptional cascade - biologically consistent results validates method
- unexpected results suggests usefulness for
system design
37Discussion
- Scalability is the issue for global robustness
analysis - curse of dimensionality exponential increase of
computational time and space - representation problem huge matrices do not
provide much intuition on phenomenon of interest
38Discussion
- Scalability is the issue for global robustness
analysis - curse of dimensionality exponential increase of
computational time and space - representation problem huge matrices do not
provide much intuition on phenomenon of interest - Global sensitivity methods provide solutions on
two counts - efficient sampling techniques (e.g. Fourier
amplitude sensitivity test)? - information aggregation via use of sensitivity
indices
39Discussion
- Scalability is the issue for global robustness
analysis - curse of dimensionality exponential increase of
computational time and space - representation problem huge matrices do not
provide much intuition on phenomenon of interest - Global sensitivity methods provide solutions on
two counts - efficient sampling techniques (e.g. Fourier
amplitude sensitivity test)? - information aggregation via use of sensitivity
indices - For parameter search problem possibility to
integrate robustness infitness function
40Thank you for your attention!