Parameter Identification using δ‐Decisions for Hybrid ...
Transcript of Parameter Identification using δ‐Decisions for Hybrid ...
![Page 1: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/1.jpg)
Parameter Identification using δ‐Decisions for Hybrid Systems in Biology
Bing Liu Joint work with
Soonho Kong, Sean Gao, Ed Clarke
CMACS/AVACS Workshop
![Page 2: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/2.jpg)
2
Biological System
Bolinsky et al. SIGGRAPH, 2006
![Page 3: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/3.jpg)
3
Computational Modeling Building a model
Structure Calibration (parameter estimation) Validation
Performing analysis
3
T imeT imeT ime
![Page 4: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/4.jpg)
4
Computational Modeling Building a model
Structure Calibration (parameter estimation) Validation
Performing analysis
![Page 5: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/5.jpg)
5
Computational Modeling Formalisms
Differential equations Boolean network Petri nets Rule‐based models … …
Multi‐mode?
![Page 6: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/6.jpg)
Activation of different networks at different stages E.g. Cell cycle, Cell differentiation
6
Hybrid Systems
Cell Cycle
Benes et al, PNAS 2009
![Page 7: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/7.jpg)
More examples: E. coli chemotaxis, kinesin walking, gene transcription, drug treatment, … …
7
Hybrid Systems
![Page 8: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/8.jpg)
Evolve in a continuous way in each mode Ruled by discrete transitions
Hybrid automata
8
Hybrid Systems
22
11
0
vdtdx
vwdtdx
q
−=
−=22 rx ≤
22
11
1
vdtdx
vwdtdx
q
−=
−=
11 rx ≤
1v
1r
1x
2v
2r
2x
w
H = !X,Q, flow,guard, reset, inv, init"
![Page 9: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/9.jpg)
A continuous space and a finite set of modes : initial configurations : continuous flows
Each mode q is equipped with differential equations
: discrete
The systems can be switched from to , resetting modes and variables
9
Hybrid Systems H = !X,Q, Init, Flow, Jump"
X ! Rk Q
Init !Q" X
Flow
Jump
d!x
dt= fq (
!x, t)
![Page 10: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/10.jpg)
How to answer questions such as: Which model structure is better? What conditions may lead to a desired state? How to control the system to avoid bad states? … …
Parameter synthesis problem Analyzing nonlinear hybrid automata is challenging
even simple reachability questions can be undecidable
10
Hurdles
![Page 11: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/11.jpg)
Use δ‐complete decision procedures to tackle the parameter synthesis problem for nonlinear hybrid models Encode a parameter synthesis problem as a first‐order
formula over the reals Perform bounded model checking Employ an interval constrains propagation (ICP) based
algorithm to identify the resulting parameters
11
Our Approach
![Page 12: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/12.jpg)
A decision procedure using numerical techniques (with an error bound δ): ϕ is false ϕ ‐δ is true
The delta‐decision problem is decidable for bounded first‐order formulas over arbitrary Type 2 computable functions exp, sin, etc, and Lipschitz‐continuous ODEs
12
Delta‐Decisions
![Page 13: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/13.jpg)
Practical tools: dReal and dReach DPLL(T), interval arithmetic, constraint solving, reliable integration, etc.
13
δ‐Complete Bounded Model Checking
![Page 14: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/14.jpg)
Prostate cancer Second leading cause of cancer‐related deaths among men in US
Hormone therapy Androgen deprivation Continuous androgen suppression (CAS) Intermittent androgen suppression (IAS)
14
Case Studies ‐ I
![Page 15: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/15.jpg)
Side effects: anemia, osteoporosis, impotence, etc. Relapse after a median duration of 18‐24 months, due to
the proliferation of androgen independent (AI) cancer cells.
15
Continuous Androgen Suppression
Time
Tumor size
CAS
![Page 16: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/16.jpg)
Reduce side effects May delay the time to relapse
Avoid emergence of AI cells
16
Intermittent Androgen Suppression
Time
Tumor size
CAS
IAS
![Page 17: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/17.jpg)
Clinical phase II and III trials confirm its advantage in terms of quality of life and cost
For time to relapse and cancer‐specific survival, its advantage depends on individual patients and the treatment scheme
How to design a personalized treatment scheme for each individual patient?
17
Intermittent Androgen Suppression
![Page 18: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/18.jpg)
Population of AD cells, AI cells, serum androgen concentration, PSA level (Ideta et al. 2008)
18
Model
yzzdx
zzm
xzzm
kzzkk
kzzkk
dtdv
zdtdz
yzzdx
zzm
dtdy
xzzm
kzzkk
kzzkk
dtdx
yy
xx
yy
xx
−
−+
−+
−−
+
−+−
+
−+=
−=
−
−+
−=
−−
+
−+−
+
−+=
βα
βα
τ
βα
βα
001
01
4
33
2
11
001
01
4
33
2
11
1
11
1)1()1(
11
1)1()1(
:flow
yzzdx
zzm
xzzm
kzzkk
kzzkk
dtdv
zzdtdz
yzzdx
zzm
dtdy
xzzm
kzzkk
kzzkk
dtdx
yy
xx
yy
xx
−
−+
−+
−−
+
−+−
+
−+=
−=
−
−+
−=
−−
+
−+−
+
−+=
βα
βα
τ
βα
βα
001
01
4
33
2
11
0
001
01
4
33
2
11
2
11
1)1()1(
11
1)1()1(
:flow
0
:jump
0
21
<+∧≤+
→
dtdy
dtdxryx
0
:jump
1
12
>+∧≥+
→
dtdy
dtdxryx
Mode 1 (on-treatment) Mode 2 (off-treatment)
![Page 19: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/19.jpg)
Proliferation, apoptosis, mutation rates Treatment thresholds: r0 and r1
19
Model
yzzdx
zzm
xzzm
kzzkk
kzzkk
dtdv
zdtdz
yzzdx
zzm
dtdy
xzzm
kzzkk
kzzkk
dtdx
yy
xx
yy
xx
−
−+
−+
−−
+
−+−
+
−+=
−=
−
−+
−=
−−
+
−+−
+
−+=
βα
βα
τ
βα
βα
001
01
4
33
2
11
001
01
4
33
2
11
1
11
1)1()1(
11
1)1()1(
:flow
yzzdx
zzm
xzzm
kzzkk
kzzkk
dtdv
zzdtdz
yzzdx
zzm
dtdy
xzzm
kzzkk
kzzkk
dtdx
yy
xx
yy
xx
−
−+
−+
−−
+
−+−
+
−+=
−=
−
−+
−=
−−
+
−+−
+
−+=
βα
βα
τ
βα
βα
001
01
4
33
2
11
0
001
01
4
33
2
11
2
11
1)1()1(
11
1)1()1(
:flow
0
:jump
0
21
<+∧≤+
→
dtdy
dtdxryx
0
:jump
1
12
>+∧≥+
→
dtdy
dtdxryx
Mode 1 (on-treatment) Mode 2 (off-treatment)
![Page 20: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/20.jpg)
Hypothesis 1 AI cells grow at the constant rate independent of the
androgen level
Hypothesis 2 AI cells do not grow when the androgen level is normal
Hypothesis 3 AI cells decrease when the androgen level is normal
Model Selection
![Page 21: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/21.jpg)
Observation When r0 = 4 (ng ml‐1) and r1 = 10(ng ml‐1), cancer relapse can
be avoided within 1000 days Define cancer relapse: PSA level > 30 ng ml‐1
Bounded model checking Property: 900≤tau≤1000 ∧v≤30 H1: unsat H2: unsat H3: sat
Model Selection
![Page 22: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/22.jpg)
Personalized parameters αy ‐ the proliferation rate of AI cells βy ‐ the apoptosis rate of AI cells m1 – the mutation rate from AD to AI cells z(0) – the initial androgen level
22
Personalized Therapy Design
![Page 23: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/23.jpg)
Different patients may response differently to the same treatment scheme (r0 = 4 and r1 = 10)
23
Personalized Therapy Design
Patient#1
Patient#2
Patient#3
Patient#3
avoided
delayed
relapse
delayed
(r0 = 4 and r1 = 10)
(r0 = 4 and r1 = 10)
(r0 = 4 and r1 = 10)
(r0 = 5 and r1 = 16)
![Page 24: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/24.jpg)
Apply IAS therapy to a patient for 1‐2 cycles and measure PSA time serials
Estimate personalized parameters by fitting PSA data Given r0 in [0,8) and r1 in [8,15], verify if H3 can reach
900 ≤ tau ≤ 1000 ∧v≤30 (i.e. no relapse) False: androgen suppression does not work True: feasible values for r0 and r1 will be returned.
24
Personalized Therapy Design
![Page 25: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/25.jpg)
Patients dataset: 109 patients Phase II clinical trials (Bruchovsky et al, Cancer, 2007)
Partial results:
25
Personalized Therapy Design
![Page 26: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/26.jpg)
Cardiac cell Minimal Resistor Model (Fenton et al, J Theor Biol, 2008) The electrical activity are governed by opening and closing of ion
channels
26
Case Studies ‐ II
![Page 27: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/27.jpg)
27
![Page 28: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/28.jpg)
When the cardiac cell loses its excitability, it will lead to disorders such as ventricular tachycardia and fibrillation
28
Cardiac Disorders
http://thevirtualheart.org/
![Page 29: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/29.jpg)
When the cardiac cell loses its excitability, it will lead to disorders such as ventricular tachycardia and fibrillation
Under what circumstances, a cardiac cell will fail to generate a successful AP (i.e. globally u < θv)? Related parameter ranges identified by Grosu et al. CAV 2011
(linear approximation) We answer this by identifying the ranges for parameters in
the original nonlinear model
29
Cardiac Disorders
!o1! (0, 0.006)"!
o2! (0, 0.13)"6.2!
o2+!
o1# 9.9
![Page 30: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/30.jpg)
30
![Page 31: Parameter Identification using δ‐Decisions for Hybrid ...](https://reader034.fdocument.pub/reader034/viewer/2022050211/626e18582c59681e4c0c66e8/html5/thumbnails/31.jpg)
A delta‐decision based framework for parameter identification Model selection Therapy optimization Diseased‐related conditions identification
What’s next More analysis for cardiac cell model Parameter estimation DNA damage induced cellular senescence
31
Summary