mathematical models of complexity - inriavideos.rennes.inria.fr/hdr/radulescu/hdr_ameliore1.pdf ·...
TRANSCRIPT
Mathematical models of complexity
Ovidiu Radulescu, IRMAR and Symbiose project IRISA
2
SummaryCV
Brief state of the art: complex systems, systems biology
Contributions in biology:
Markov processes in molecular biology
Qualitative equations for functional genomics
PDE models for pattern formation
Conclusion
3
CVEducation: 1989 Diplôme d’Ingénieur Physique des Solides, Bucarest1994 Doctorat Physique des Solides, Orsay (félicitations)1996 DEA probabilités, Marne-la-Vallée
Recherche: interdisciplinarité, transversalité2 post-docs (Pays Bas et Angleterre) 27 articles acceptés, 12 proceedings conf.
Enseignement:1991-1993, moniteur physique Orsay, vacataire Ecole Centrale de Paris1993-1996 ATER et PRAG physique, Marne la Valléedepuis 1999 MC en mathématiques à Rennes 1encadrement d’une thése (en mathématiques) et d’une dizaine de stages
Responsabilités:membre commission informatique, animation d’un groupe de travailcoordinateur d’une ACI
4
Complex systems
Quasicrystals(Orsay)
Cellular physiology(Rennes) Development
(Rennes)
Incommensurate composites(Nimégue)
Wormlike micelles(Leeds)
5
What complex systems have in common
• Order as framework for transformation: crystals, dissipative structures, patterns
• Defects as motors for transformation: points, lines, interfaces
• Hierarchical organisation
• Nonlinearity
• Stability, robustness
• Universality
6
Systems biology
• Mathematical modeling of physiology
• Transversal field, imports methods from physics, control theory, automata, chemical kinetics
• After rapid evolution, critical stage: obstacle raised by the complexity of higher organisms (models are scarce or weakly predictive)
• There is a need for new methodsanalysis methods for massive datamodel reductionmore realistic models using physico-chemistry
7
Gene regulation is the result of many interactions
8Network models unify various processes
Gap genes, first 3 hours of DrosophilaLactose operon, E.Coli
Nutritional stress, E.Coli
9
Chromatin ImmunoPrecipitation on Chip
DNA Chip
Various kind of data: differences of concentrations, direct test of qualitative interaction
10
StrategyAims:• Model construction• Model analysis• Biological predictions
Difficulties:• Data collection is massive but unguided• Reverse engineering is difficult• Models are non-linear and in very high dimension• Interpretation of computer simulations is difficult
My solutions:• Guide data collection (experiment design)• Do not start reverse engineering from scratch (model correction)• Develop new mathematical techniques for model analysis• Look for network design principles
11
Jump Markov processes
Ordinary differential equations
Qualitative equations
Partial differential equations
Thermodynamic limitPiecewise deterministic
Discretisation
Partial thermodynamic limit
Averaging
My mathematical garden
12
My contributions
My contributions to this field:1) Modeling stochasticity of molecular biology processes by
piecewise deterministic Markov processes2) Qualitative equations for analysis of massive data3) Carr-Pego type model reduction for pattern formation4) Measure concentration as framework for robustness
CollaborationsComputer scientists: A.Siegel, M.LeBorgne (IRISA Symbiose),
M.Samsonova(St.Petersburg)Biologists: N.Theret (INSERM), S.Lagarrigue (INRA), A.Lilienbaum
(CNRS), J.Reinitz (Stony Brook)Mathematicians: S.Vakulenko(St.Petersburg), A.Gorban(Leicester),
E.Pécou(Nice)
Research project MathResoGen (2003-2006)
Modeling stochasticity in molecular biology by Markov processes
14
Markov jump processes: Renyi, Bartholomay, 50’
Modeling stochastic effects
are n chemical species
rninZiii ,1, =∈−= αβθ
nininini AAAA ββαα ++++ ←→ ...... 1111
nAA ,...,1
biochemical reaction
nZX ∈ is the state
∑=
−−+ +=nr
iXiXi
iiXqXqX
1(.)])((.))([,.)( θθ δδμ
∑=
−+=nr
iii XVXVX
1)]()([)(λ
∑=
−+=nr
jjjii XVXVXVXq
1)]()([)/()(
intensity
distribution of jumps
jump probability
jump vector
15
Suppose that the mass action law is satisfied
∑=
==nr
iii xxFsxFds
sdx1
)(v)()),(()( θ
Thermodynamic (deterministic) limit
Rescale the process ΩiXix /=
∏=
=Ω=n
s
issiiii XXXV
1xk)(v),(v)(
α
∏=
−−−− =Ω=n
s
issiiii XXXV
1xk)(v),(v)(
β
For the Markov jump processes converges in probability to the solution of a system of ordinary differential equations (Kurtz, 70)
∞→Ω ix
Ω : reaction volume
16
Some species are in small numbers!
Piecewise deterministic limit
0 , ),,()(~
≠∀ΩΩ= iri
XXVXV
rf
ii θε
,∞→Ω
O.Radulescu, A.Muller, A.Crudu (TSI in press)
,0→ε
use frequent/rare species decomposition
),( rXfXX =
1→Ωε
0 , ),,(v)(~
=∀ΩΩΩ= iri
XXXV
rf
ii θε
mass action law is not applicable and should be replaced by
concentration of rare species
reactions acting on rare species
reactions not acting on rare species
17
Piecewise deterministic limit result
For the Markov jump process converges to a piecewise deterministic process:
1,0, →→∞→ εε ΩΩ ),/( rf XXX Ω=
O.Radulescu, A.Muller, A.Crudu (TSI in press)
i0i
i
~
v))(),(()(
θθ∑
=
==r
ff
srXsxfFdssdx
)(sx fis discrete and jumps with intensity
Between two jumps is continuous and satisfies:
)(srX ),(~
rfi XxV
18
Application: hybrid stochastic simulation algorithm
O.Radulescu, A.Muller, A.Crudu (TSI in press)
0,, 00 === tXXxx rrff
)](exp[~ , rXx fλτ)()( τ+→ txtx ff
rXτ+→tt
1. Initialize2. Generate exponential random time3. Use deterministic solver to propagate4. Change to a new discrete value5. Increment time6. If t<tmax goto 2
19
Application to happloinsufficiency
Biological problem:
Syndrom due to deficient genotype : insufficient copy numberPhenotype: heterogenous cell populations
Aim:
Find the simplest model that reproduces this situation
20
Model for haploinsufficiency
dP/dt= - k3PdP/dt=k2-k3P
k-1 k1
Markov jump model (Cook 99)
*1
1GG
k
k−←→
PGGk
+→ ** 2
3kP→
Result: If the model allows a piecewise deterministic approximation
)(2 Ω=Ok
⎪⎩
⎪⎨⎧
=−
=+−=
0if,1if,
*3
*23
GPkGkPk
dtdP
O.Radulescu, A.Muller, A.Crudu (TSI in press)
Study intermittency of trajectories and the invariant distribution
1*=+GG
Ω= /1ε
21
Conclusion
Results:• The protein production is intermittent• The heterogeneity of the phenotype can be described by a Beta
distribution
The same method will be applied to larger, more complex models; in project NFκB signaling
Qualitative equations
23
Qualitative equationsBiological problem:
Following a perturbation (stress, signal) the state of the cell changes. Variations of hundreds or thousands of variables can be monitored. How to use this information?
Steps:
• develop an “elasticity” theory of graphs (O.Radulescu et al. J.R.Soc.Interface 2006)
• translate this theory into qualitative equations (with A.Siegel et al. Biosystems 2006)
• polynomial algorithms for solving systems of qualitative equations (with Ph.Veber, M.leBorgne et al. Complexus 2006)
• application to huge networks (with C.Vargas et al., proc. RIAMS 2006)
24
Elasticity of graphs),( PXFdt
dX = 0),( =PXFSteady state is perturbed
∑∑→ →
→
∂∈=
ijj
ij
ij
Gji X
C
aX δδ
G∂ Perturbations propagate along pathsin the directed graph
0 iff arcan is ),( ≠∂
∂=
i
j
X
Fjiaij
XP δδ →
graph boundary
(O.Radulescu et al. J.R.Soc.Interface 2006)
Steady state equationdynamics
Dirichlet solution:
25
Qualitative equations
Li=Le+LacY-LacZ
∑∈∂
∂ −⎟⎠
⎞⎜⎝
⎛−=
)(
1ipredj
jjiiX
iF
i XaX δδ
subgraph
∑∈
=)(
)) ()((ipredj
jjii XsignasignXsign δδ
?,,
Dirichlet solution for subgraph
Qualitative equation
+−∈sign
(with A.Siegel et al.Biosystems 2006, with Ph.Veber, M.LeBorgne Complexus 2006)
Sign algebra
26
Algorithm for solving qualitative equations•Map signs to elements of the finite field Z/3Z
•Map qualitative equations to polynomial equations over Z/3Z
•NP complete problem
•Ternary Decision Trees contracted to directed acyclic graphs andsystematic use of cache memory for non-redundant computation
•Obtain exhaustive lists of solutions within minutes for 1000 nodes
(with Ph.Veber, M.LeBorgne Complexus 2006)
27
Predictions of a model
hard components: variables whose values are the same (+ or -) in any solution the hard components are the predictions of the model
Le, cAMP, A are observed
Li,G,LacZ,LacY, LacI are hard components
28
Experiment design
Any value of the triplet(Le,G,A) can be extendedto a solution
These variables have no validation power
29
Use validation power for experiment design
Only 2 values (out of 8) of (LacI,A,LacZ), namely(+,−, −) (−, +,+) can be extendedto a solution
Define validation power as:
Choose high validation power sets for optimal design
30
Large scale application: nutritional stress of E.Coli1258 nodes, 2526 interactions, 10600 states, 1016 solutions
We have obtained both:• a set of predictions: from 40 observations in the stationary phase,
401 hard components, 26% of the network• a set of corrections to the model: necessarily include σ factors
Partial differential equations
32
Pattern formationProblem:• Patterns form in very different complex systems (Drosophila
embryo before gastrulation, shear banding of complex fluids). • The examples are of Wolpert type, less studied in mathematics.
Can we find an unified approach?
Cornerstones:• of complex fluids: understand the relation between structure and
flow properties• of developmental biology: understand canalization, stability of
development
Collaborations:P.D.Olmsted (Physics,Leeds), JP.Decruppe(Physics,Metz), JF.Berret, G.Porte
(Physics,Montpellier) on wormlike micellesS.Vakulenko (Maths,St.Petersburg), J.Reinitz(Appl.Maths and Biology, Stony Brook) on
Drosophila
33
Problem 1:
Syncitial blastoderm, before gastrulation
34
Synt
hesi
sTr
ansp
ort
Dec
ay
∑=
++=∂∂ N
baababaa
a hxmTtxuTgRttxu
1))(),((),(
),(2 txuD aa∇+
),( txuaaλ−
Model Reaction-diffusion equations
35
Synt
hesi
s
Genetic Interconnectivity Matrix (T):
T parameters:
positive: negative: zero:
activationrepressionno interaction
1
2
…
N
1 2 … NabGene
T11 T12 …
…
…
T22T21
TN1 TN2
… …
T1N
T2N
TNN
…
∑=
++=N
baababaa
a hxmTtxuTgRdttxdu
1))(),((),(
36
Synt
hesi
s
Action of maternal gradient (bicoid)
Bicoid profile m(x) develops in 1h after fertilizationand remains constant during the blastoderm
∑=
++=N
baababaa
a hxmTtxuTgRdttxdu
1))(),((),(
37
Synt
hesi
sTr
ansp
ort
Dec
ay
+Da (n) vi−1a − vi
a( )− vi+1a − vi
a( )[ ]
−λavia
The regulation-expression function g(u):1.0
0.0-6.0 0.0 6.0
Rel
ativ
e A
ctiv
atio
n: g
(u)
Total Regulatory Input: u
∑=
++=N
baababaa
a hxmTtxuTgRdttxdu
1))(),((),(
38
Problem 2: Shear banding of wormlike micellesHadamard instability
O.Radulescu et al. Rheol.Acta 1999, with PD.Olmsted J.Rheol. 1999
39
Model: Fluid-structure coupling
σρ .v)v.( ∇=∇+∂t
Navier-Stokes
ττμ //2)()()v.( 2 Σ−Δ+Σ∇=ΣΔ+ΔΣ−ΣΩ−ΩΣ−Σ∇+∂ Dat
Johnson-Segalman constitutive model + stress diffusion
Re=0 approximation
.,0..constS =+==∇ γεσσ )1(
.
2
2
WSyS
DtS −+−
∂
∂=∂
∂ γτ
SWyW
DtW .
2
2
γτ +−∂
∂=∂
∂
principal flow equations
Stress dynamics is described by a reaction-diffusion system
O.Radulescu, PDOlmsted J.Non-Newtonian.Fl.Mech. 2000
40
Common framework: R-D PDE with small diffusion
),,(22 txufuDut εε +∇=
Ω⊂Ω∈ ,qRxnRtxuu ∈= ),(
dndddiagD ,...,2,1=
)()0,( 0 xuxu =
Ω∂∈=∇ xxnxu ,0)().(
Cauchy problem for the PDE system
is compact with smooth frontier
initial data
no flux boundary conditions
idea : consider the following shorted equation
),,v(v txft ε=
41
Result 1: Classification of patterning mechanisms
0when0,t,in uniformly ,0t)-v(x,t)(x,u →>Ω∈→ εε x
t)(x,uε solution of the full system
v(x,t) solution of the shorted equation
Patterning is diffusion neutral if for vanishing diffusion, the solution of the full system converges uniformly to the solution of the shorted equation
If not, patterning is diffusion dependent
O.Radulescu and S.Vakulenko, arXiv qBio 2006
42
Result 2: Classification of interfaces
For a given x, the shorted equation has only one attractor
Type 1 interface
O.Radulescu and S.Vakulenko, arXiv qBio 2006
Type 2 interface
For a given x, the shorted equation has severalattractors, here 2:
(x)φ
(x)(x), 21 φφ
Patterning with type 1 interfaces is diffusion neutral
Patterning with type 2 interfaces is diffusion dependent
The width of type 2 interfaces can be arbitrarily small
43
Theorem on the diffusion neutral patterning
O.Radulescu and S.Vakulenko, arXiv qBio 2006
Consider the time autonomous situation and the shorted equation
The patterning is diffusion neutral under the following conditions on the shortedequation:
i) uniform dissipativity
ii) strong linear stability
calculated at the attractor
iii) attraction basin condition
44
Theorem on the movement of type II interfaces in the bistable case
O.Radulescu and S.Vakulenko, arXiv qBio 2006
Invariant manifold decomposition for
Travelling wave solution for the space homogeneous eq.
The solution of space inhomogeneous equation is of the moving interface type
Equation for the position q(t) of the interface
This extends results of Carr-Pego(90) and Fife (89)
The velocity of a Type II interface is proportionalto the square root of the diffusion coefficient
45
Application1: stress diffusion coefficient from interface kinetics
Diffusion is smallD~0.003-0.011 μm2s-1
w~30-40 nm
O.Radulescu et al. Europhys.Lett. 2003
10s-130s-1 .γ
60s-1
σ.γ
46
Application2: Diffusion dependent patterning of Drosophila
1) parameter fit of Reinitz model from time dependent data by simulated annealing
2) compute attractors of shorted equation
Result: patterning is diffusion dependentImprovement of model fit: Rapid method of parameter identification using
interface kinetics
with Gursky, Manu, Vakulenko, unpublished
presented at Nanobio’06, St.Petersburg
47
Comment on the impact in biology
Compared to Turing models, the gene circuit model is realistic:• the pattern is not a periodic modulation of a homogeneous state• the pattern results from the interaction of development genes, is
guided by maternal gradients and has aperiodic transients
Treating the set of segmentation genes as a dynamical system allows to understand:
• The logic of interactions (open problem) and transformations • The stability of the result (open problem)• The possible errors in mutants (open problem)
48
Conclusion and future projectsStart of a long term project: produce powerful mathematical tools for analysis of complex systems.
Strategy: Model simplification *The invariant manifold technique of Carr-Pego*piecewise deterministic approach for Markov processes *graph theory methods for chemical kinetics models
An intrinsic relation exists between model reduction, stochasticityand robustness: concentration phenomena!
Physical chemistry for diffusion and transport in physiology.
CollaborationUpi Bhalla NCBS Bangalore, planned co-tutored phD.A.Gorban (Leicester) Egide/Alliance sponsorship J.Reinitz (Stony Brook) and Samsonova (St.Petersburg)ASC project with INRA on modeling lipid metabolismproject ANR SITCON with CurieSymbiose team IRISA
49
Project 3: Robustness of biological systems
Cube concentration
Distributed robustness
r-robustness
Simplex concentration
Stony Brook (Reinitz), Leicester (Gorban), St.Petersburg(Samsonova, Vakulenko)
50
Project 4: LIPID METABOLISM
Hierachical modeling:1)Extended model2) Abstract model
Multiorgans, multispecies
Heterogeneous dataMicroarraysBiochemical dosages
Symbiose, INRA Rennes and Toulouse
51
Project 5: SITCONModeling signal transduction induced by a
chimeric oncogene
CELL PROLIFERATIO
N
Νfκb pathway
TGFβ pathway
IGF pathway
…
EWS/FLI1
Ewing network
APOPTOSIS
EWS/FLI1Institut Curie, Symbiose
Farey sets and spectra of incommensurate structures
53
Number theory and Incommensurate compounds
Related problems:Hofstadter Butterfly
Bellissard’s gap labellingArnold tongues