Approximation of a cell cycle model in Budding Yeast
• Introduction • Cell Cycle Model of Budding Yeast • Constant Input Response (CIR) of the Full Model • Proposed Method • Preprocessing: Jacobianbased Local Refinement (JLR) • Initialization of the Genetic Algorithm (GA)
• Results and Discussions Comparison between 2 Scenarios for Substitution of Excluded States
•
Results of Substitution Excluded States with Mean Values • GA with Parameter Estimation
•
Results of the GA with Parameter Estimation • Conclusion and Future Works
•
Presented by Chaiyut Thanukaew
Introduction
• Molecular mechanisms in biological systems are complex and have dynamic interactions. • To express behaviors of a system, often a set of nonlinear ODEs is occupied. • Since normally huge models, mathematical model reduction is used to gain insight into the models and reduce computational cost.
• There are 36 ODEs, 25 algebraic equations and 139 parameters in the cell cycle model of budding yeast.
• The reduced model must retain essential features of the full model, traditionally, the trajectories of certain state variables. Moreover, it must comply with viability criteria.
• For the cell cycle model of budding yeast, responses of the mass regarding external input is decided to be the preserved characteristic .
of the model by using genetic algorithm.
Cell Cycle Model of Budding Yeast (1)
S
p
indle
Nuclear
Formation
Migration
4 phases of eukaryotic* cell cycle; G1, S, G2
and Mphase.
• G1 or Gap1 phase: the cell grows.
G2
Chromosome
M
Segregation,
• S or DNA synthesis: the cell makes
copies of its chromosomes and get ready
to divide.
DNA
Nuclear
S
Replication
Division
• G2 or Gap2 phase: the cell check the
duplicated chromosomes.
Bud
G1
Emergence
Cytokinesis
(Cell Division)
• M or Mitosis phase: the cell separates the
copies chromosomes into 2 full sets and
the cell divides into 2 new cells
Spindle
Pole, Body
Duplication
Start
Growth
How to know which mechanisms of the cell cycle
take place at what time?
CDKCyclin** is a protein active in regulating
Figure 1: Cell Cycle of Budding Yeast
the cell cycle.
* Eukaryote is a cell with visible nucleus e.g. yeast and mammal. ** Cyclin Dependent Kinase and Cyclin
Cell Cycle Model of Budding Yeast (2)
Viability Criteria • Correct sequences, i.e.
Budding DNA synthesis
segregation Cell division. • [mass] <= 10 all time.
Figure 2: Consensus Model of Budding Yeast
Constant Input Response (CIR) of Full Model (1)
CIR definition: “For each added constant input, the CIR is the amplitude and time of the
mass state such that it becomes stable while, also, the whole model must comply the
viability criteria”.
Constant In ut added to the state 5 (Sic1)
p
;
d Sic
[
1]
dt
(
k
'
k
"
.[
Swi
5]) (
V
k
).[ 2] (
C
V
k
).[ 5]
C
k
.[
Cdc
14].[ 1 ]
Sic P
k
Clb
2
k
Clb
5
] Vkp, c1).[
Si
c 1 ]
K
Why adding at [Sic1]?
a constant signal, physically, would mean putting some substances to boost up some
activities of the cell cycle, e.g. going faster into the G1phase.
How to notice the response of the budding yeast model after adding up the constant
signal to the Sic1 state?
perturbation.
Constant Input Response (CIR) of Full Model (2)
Response of the mass to different constant inputs
Figure 3: The mass state when different constant inputs are applied
Proposed Method (1)
The full model is written as;
x f ( x , p )
We rewrite the full model as;
x f ( x , s , p )
Where;
x: the vector of states
p: the vector of parameters, and
s: a vector where s _{i} is “1” when the ith state is included in the model
(“0” if excluded).
Objective
The reduced model
x f x s p
(
,
,
)
is defined such that its solution x
minimizes
the cost function
, i.e.
( x ) min
( x )
Where;
: the cost function, and
S : the set of all vectors of length N _{f} (Number of states of the full model)
whose entries are “0” or “1”.
Proposed Method (2)
Cost Function
The cost function
is defined as;
_{}
_{} _{1}
1. 1.
X is not oscillatory or inviable.
X is oscillatory and viable.
Where;
The relative model size _{} is,
N
N
the number of states in the reduced model
the number of states in the full model
Proposed Method (3)
Where (continued);
The CIR error is computed as the least square distance between one of the
full and one of the reduced model as following;
N
per
2
2
(
x
( )
i
x
( ))
i
(
y
( )
i
y
( ))
i
F
R
F
R
i 1
N
per
2
2
x
( )
i
y
( )
i
F
F
• x _{F} and y _{F} represent the positions of the CIR of the full model when
different constant inputs are applied,
• N _{p}_{e}_{r} is the maximum number of different constant inputs for which the
full model is still oscillatory and viable, and
• x _{R} and y _{R} represent the positions of the CIR for the reduced model
regarding different constant inputs added.
Preprocessing : JLR (1)
Jacobian matrix and JLR
relationship between equations in a system regarding that vectorvalued function.
Considering the following equation;
J J diag ( J )
Where,
J: the Jacobian matrix, and
diag(J): all diagonal elements of the Jacobian matrix.
If there is any row in
J
such that all elements are zeros inputonly state, and


If there is any column in

J

such that all elements are zeros outputonly state.

only state can be substituted by a constant and the outputonly state can be indirectly
obtained by some other states.
Preprocessing : JLR (2)
Index table of the original cell cycle model
Inputonly
states
~
Table 1: Index Table
J J diag ( J )
of the Cell Cycle Model
Outputonly
12
states
Initialization of the GA (1)
Problem: Very hard to obtain individuals being oscillatory and viable in the gen. “0”.
Reasons would be as following;
•

Generating individuals by rounding a random number of uniform distribution

causing that half of the number of bits is “0” in average (18 from 36 bits). There is

the problem, e.g. assume that oscillatory individuals must comprise of at least 26

states, thus it is hard to obtain this number.

•

The generated individuals lack of some crucial states, e.g. mass and some states of

cyclins. Thus, leaving out only one of these states causes the reduced models non
oscillatory or inviable.
Solution introduce preconditions to generates the individuals in the initial generation
of the GA.
Initialization of the GA (3)
Example: Feasible individual generated in gen. “0” with preconditions
bit 1
bit 2
bit
3
bit 4
bit
5
bit 26
bi t 27
bi t
33
bit
34
bit
35
bit
36
1
R
R

State 1: set to “1” (mass)

State 3 and 4: randomly generated

State 4: set to “1” (viability concerned)

State 5: set to “1” (constant input added)
1
State 33, 34, 35 and 36: set to “1s”
(viability concerned)
Figure 5: Example of Generating a Feasible Individual in Gen. “0”
Comparison between Zero VS Meanvalued Sub.
2 scenarios of substitution an excluded state;
• Substitution with zero: the excluded states is completely cut away, and
• Substitution with mean value: the excluded state still has influence, but not
dynamic behavior.





























Whole population replacement with elitist

Whole population replacement with elitist

Whole population replacement with elitist




Table 2: Parameter Setting of 2 Substitution Scenarios
Results : Comparison between Zero VS Meanvalued Sub.
The best result regarding Table 3 comes from the meanvalued substitution

2. Reduced model with meanvalued substitution












Table 3: Results of Meanvalued Substitution
simulations since it returns the better cost value (the lower value).
outputonly states added to determine the viability (see slide 12).
scenario even population size and crossover strategy are varied.
Results : Meanvalued Substitution (1)
Table 4: Results of Meanvalued
Substitution from Several Simulations
Purpose: to prove that the results
do not come from local minima .
• The best reduced models “S1” and
“S2” contains 23 states.
• Although, different number of
states for each result, all results are
somehow in the same pattern (blue
highlight).
• If fixing highlight areas, the best
result can be obtained from varying
of 10 bits in nonhighlighted areas
(search space is reduced).
• Further test “S1” and “S2”
comparing their responses to one of
the full model .
Results : Meanvalued Substitution (2)
The cell cycle of the full model VS the cell cycle of the reduced model
S1
in Table 4
Figure 6: Cell Cycle of the Full Model VS Cell Cycle of
Results : Meanvalued Substitution (4)
Reduced model “S1” as the Jacobian index table
: Excluded
states (non
dinamic)
: Substituted
by constant
: Nonzero
Jacobian
19
Table 5: Reduced Model “S1” Substituting in the Jacobian Index Table
Results : Meanvalued Substitution (4)
Reduced model “S1” as the Jacobian index table
Table 6: Jacobian Index Table of the Reduced Model “S1”
Note: State 24 (Tem1) is inputonly state but cannot be left out since it
involves resetting parameters.
GA with Parameter Estimation (1)
Idea: number of feasible solutions might be increased if the thresholds of determination
the viability criteria are changed.
Implementation method: to put 4 more bits representing real value numbers used as the
thresholds of the viability determination.


Binary bit strings and 4 positive real values










Uniform crossover with prob. = 0.5 (crossover in the parameter part is

substituted by the median values of the parents)


Bit mutation with prob. = 0.1 and normal distribution added with the

previous value in the parameter part


Whole population replacement with elitist no. = 2


Table 7: Parameter Setting of the GA with Parameter Estimation
Results: GA with Parameter Estimation
Results of the GA with Parameter Estimation

Reduced Model with

Parameter Estimation

Round#3

Round#3

Ind.1

Ind.2

Kez (0.3)



Kez2 (0.2)



Thres. 0.1



Thres. 1



Total State

24

25

No.

Total Cost

0.7765

0.7954

Discussions
The results are not better than the results with no
parameter estimation. The reasons would be as
following;
• Inappropriate estimation of “Thres. 1”
because it is used as the threshold for 3 states
[ORI], [BUD] and [SPN],
• Newly estimated parameters cannot lead
feasible individuals into optimal region in the
search space, and
• New parameters are generated by the normal
distribution with a small variance.
Table 8:The 2 Best Result of GA with Parameter
Estimation (from 3 times of simulation)
Conclusion and Further Works
• The best result contains 23 states. 3 more states can be ignored because of the
outputonly states.
• Appropriate GA’s parameters might increase efficiency of the simulation, e.g. lessen
the computational time, explore the search space intelligently.
• The GA hybridizes with the Jacobianbased local refinement can reduce
unnecessary computation, e.g. the GA with JLR in this work.
• Although GA can be applied to resolve a given problem without prior knowledge of
such that problem, initialization of the GA is applied to this problem because of
physical reality and model constraints.
• Further works would be to study more about the GA with parameter estimation and to
combine other strategies of model reduction to the GA, e.g. to lump 2 continuous states
together.
Thank You
Appendix A : CIR of the Full Model
Figure A.1: CIR of the full model such

Fi

ure A 2: CIR of the full model such


g

.

that Kadded is increased 0.01 each

that Kadded is the median values of

time (from 0 to 0.28)

each group
