Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Model Reductions by Subgraph Epimorphisms
Master MPRI C2-19
Computational Methods in Systems and Synthetic Biology
Fran¸cois Fages
Inria Saclay, Lifeware team, France
after Steven Gay’s PhD Thesis (former MPRI C2-19 student 2008-2009)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
From Models to Metamodels
In Systems Biology, models are built with two contradictory perspectives:
models for representing knowledge: the more detailed the better models for making predictions:
the more abstract the better ! get rid of useless details
These two perspectives can be reconciled by organizing models in a hierarchy of modelsrelated by reduction/refinement relations.
To understand a system is not to know everything about it, but to know abstraction levels that aresufficient
for answering given questions about it
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
From Models to Metamodels
In Systems Biology, models are built with two contradictory perspectives:
models for representing knowledge:
the more detailed the better
models for making predictions:
the more abstract the better ! get rid of useless details
These two perspectives can be reconciled by organizing models in a hierarchy of modelsrelated by reduction/refinement relations.
To understand a system is not to know everything about it, but to know abstraction levels that aresufficient
for answering given questions about it
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
From Models to Metamodels
In Systems Biology, models are built with two contradictory perspectives:
models for representing knowledge:
the more detailed the better models for making predictions:
the more abstract the better ! get rid of useless details
These two perspectives can be reconciled by organizing models in a hierarchy of modelsrelated by reduction/refinement relations.
To understand a system is not to know everything about it, but to know abstraction levels that aresufficient
for answering given questions about it
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
From Models to Metamodels
In Systems Biology, models are built with two contradictory perspectives:
models for representing knowledge:
the more detailed the better models for making predictions:
the more abstract the better ! get rid of useless details
These two perspectives can be reconciled by organizing models in a hierarchy of modelsrelated by reduction/refinement relations.
To understand a system is not to know everything about it, but to know abstraction levels that aresufficient
for answering given questions about it
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
State of the Art: Model Repositories
biomodels.net: plain list of583 manually curated and 796 non curated models in SBML format
MAPK signaling cascade
009 Huan: three-level cascade of double phosphorylations
010 Khol: reduced model without dephosphorylation catalysts 011 Levc: same model as 009 Huanwith different parameter values and different molecule names
027 Mark, 028 Mark, 029 Mark, 030 Mark, 031 Mark: reduced one-level models with different levels of details Circadian clock:
074 Lelo, 021 Lelo, 170 Weim, 171 Lelo, ...
Calcium oscillation: 122 Fish, 044 Borg, 117 Dupo,... Cell cycle: 056 Chen, 144 Calz, 007 Nova, 169 Agud,...
•Relations between molecule names may be given in annotations
•No relations between models (given in the articles at best)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
State of the Art: Model Repositories
biomodels.net: plain list of583 manually curated and 796 non curated models in SBML format
MAPK signaling cascade
009 Huan: three-level cascade of double phosphorylations 010 Khol: reduced model without dephosphorylation catalysts
011 Levc: same model as 009 Huanwith different parameter values and different molecule names
027 Mark, 028 Mark, 029 Mark, 030 Mark, 031 Mark: reduced one-level models with different levels of details Circadian clock:
074 Lelo, 021 Lelo, 170 Weim, 171 Lelo, ...
Calcium oscillation: 122 Fish, 044 Borg, 117 Dupo,... Cell cycle: 056 Chen, 144 Calz, 007 Nova, 169 Agud,...
•Relations between molecule names may be given in annotations
•No relations between models (given in the articles at best)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
State of the Art: Model Repositories
biomodels.net: plain list of583 manually curated and 796 non curated models in SBML format
MAPK signaling cascade
009 Huan: three-level cascade of double phosphorylations 010 Khol: reduced model without dephosphorylation catalysts 011 Levc: same model as 009 Huanwith different parameter values and different molecule names
027 Mark, 028 Mark, 029 Mark, 030 Mark, 031 Mark: reduced one-level models with different levels of details Circadian clock:
074 Lelo, 021 Lelo, 170 Weim, 171 Lelo, ...
Calcium oscillation: 122 Fish, 044 Borg, 117 Dupo,... Cell cycle: 056 Chen, 144 Calz, 007 Nova, 169 Agud,...
•Relations between molecule names may be given in annotations
•No relations between models (given in the articles at best)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
State of the Art: Model Repositories
biomodels.net: plain list of583 manually curated and 796 non curated models in SBML format
MAPK signaling cascade
009 Huan: three-level cascade of double phosphorylations 010 Khol: reduced model without dephosphorylation catalysts 011 Levc: same model as 009 Huanwith different parameter values and different molecule names
027 Mark, 028 Mark, 029 Mark, 030 Mark, 031 Mark:
reduced one-level models with different levels of details
Circadian clock:
074 Lelo, 021 Lelo, 170 Weim, 171 Lelo, ...
Calcium oscillation: 122 Fish, 044 Borg, 117 Dupo,... Cell cycle: 056 Chen, 144 Calz, 007 Nova, 169 Agud,...
•Relations between molecule names may be given in annotations
•No relations between models (given in the articles at best)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
State of the Art: Model Repositories
biomodels.net: plain list of583 manually curated and 796 non curated models in SBML format
MAPK signaling cascade
009 Huan: three-level cascade of double phosphorylations 010 Khol: reduced model without dephosphorylation catalysts 011 Levc: same model as 009 Huanwith different parameter values and different molecule names
027 Mark, 028 Mark, 029 Mark, 030 Mark, 031 Mark:
reduced one-level models with different levels of details Circadian clock:
074 Lelo, 021 Lelo, 170 Weim, 171 Lelo, ...
Calcium oscillation: 122 Fish, 044 Borg, 117 Dupo,...
Cell cycle: 056 Chen, 144 Calz, 007 Nova, 169 Agud,...
•Relations between molecule names may be given in annotations
•No relations between models (given in the articles at best)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
State of the Art: Model Repositories
biomodels.net: plain list of583 manually curated and 796 non curated models in SBML format
MAPK signaling cascade
009 Huan: three-level cascade of double phosphorylations 010 Khol: reduced model without dephosphorylation catalysts 011 Levc: same model as 009 Huanwith different parameter values and different molecule names
027 Mark, 028 Mark, 029 Mark, 030 Mark, 031 Mark:
reduced one-level models with different levels of details Circadian clock:
074 Lelo, 021 Lelo, 170 Weim, 171 Lelo, ...
Calcium oscillation: 122 Fish, 044 Borg, 117 Dupo,...
Cell cycle: 056 Chen, 144 Calz, 007 Nova, 169 Agud,...
•Relations between molecule names may be given inannotations
•No relations between models (given in the articles at best)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Detect Reductions from the Structure of the Reactions ?
Mathematical methods based on slow/fast reactions currently do not apply on a large scale(Thesis subject using tropical algebra)
⇒Purely graphical method based on the reactant/product structureabstracting from names, kinetics and stoichiometry ?
Example (Hierarchy of MAPK models in biomodels.net computed from the structure of their reactions)
009_Huan
010_Khol 011_Levc
027_Mark
029_Mark 031_Mark
026_Mark
028_Mark 030_Mark 049_Sasa
146_Hata
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Detect Reductions from the Structure of the Reactions ?
Mathematical methods based on slow/fast reactions currently do not apply on a large scale(Thesis subject using tropical algebra)
⇒Purely graphical method based on the reactant/product structureabstracting from names, kinetics and stoichiometry ? Example (Hierarchy of MAPK models in biomodels.net computed from the structure of their reactions)
009_Huan
010_Khol 011_Levc
027_Mark
029_Mark 031_Mark
026_Mark
028_Mark 030_Mark 049_Sasa
146_Hata
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Reaction Hypergraphs (Petri net structure)
Definition
Areaction hypergraph is a bipartite graph (S,R,A) whereS is a set ofspecies,R is a set ofreactionsand A⊆S×R∪R×S.
Example (E +S ES →E+P and E+S →E+P)
E c ES
d S
p P
S c P
E
not contained in the previous graph no subgraph isomorphism
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Model Reductions by Graph Operations
In our setting, a model reduction is afinite sequenceof graph reduction operations of four types:
1 Species deletion
deletion of one species vertex with all its incoming/outgoing arcs
2 Reaction deletion idem
3 Species merging
replacement of two species vertices by one species vertex with all their incoming/outgoing arcs
4 Reactions merging idem
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Model Reductions by Graph Operations
In our setting, a model reduction is afinite sequenceof graph reduction operations of four types:
1 Species deletion
deletion of one species vertex with all its incoming/outgoing arcs
2 Reaction deletion idem
3 Species merging
replacement of two species vertices by one species vertex with all their incoming/outgoing arcs
4 Reactions merging idem
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Model Reductions by Graph Operations
In our setting, a model reduction is afinite sequenceof graph reduction operations of four types:
1 Species deletion
deletion of one species vertex with all its incoming/outgoing arcs
2 Reaction deletion idem
3 Species merging
replacement of two species vertices by one species vertex with all their incoming/outgoing arcs
4 Reactions merging idem
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Model Reductions by Graph Operations
In our setting, a model reduction is afinite sequenceof graph reduction operations of four types:
1 Species deletion
deletion of one species vertex with all its incoming/outgoing arcs
2 Reaction deletion idem
3 Species merging
replacement of two species vertices by one species vertex with all their incoming/outgoing arcs
4 Reactions merging idem
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Example of the Michaelis-Menten Reduction
E c ES
d S
p P c+p
ES E
d S
P
c+p
ES E
d S
P c+p
ES E
S P
merge(c,p)
delete(d) delete(ES)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Example of the Michaelis-Menten Reduction
E c ES
d S
p P c+p
ES E
d S
P
c+p
ES E
d S
P c+p
ES E
S P
merge(c,p)
delete(d)
delete(ES)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Example of the Michaelis-Menten Reduction
E c ES
d S
p P c+p
ES E
d S
P
c+p
ES E
d S
P c+p
ES E
S P
merge(c,p)
delete(d) delete(ES)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Commutation Properties of Delete/Merge Operations
The merge and delete operations enjoy the following commutation and association properties:
How to detect graph reductions efficiently ? Without loosing time in symmetrical solutions ?
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Subgraph Epimorphisms
Definition
Asubgraph morphism(SMOR) fromG = (S,A) to G0= (S0,A0) is a functionµ:S0 −→S0, with S0 ⊆S such that
∀(x,y)∈A∩(S0×S0) (µ(x), µ(y))∈A0.
Asubgraph epimorphism (SEPI) is a surjective SMOR
∀x0 ∈S0 ∃x ∈S0 µ(x) =x0,
∀(x0,y0)∈A0 ∃(x,y)∈A µ(x) =x0 µ(y) =y0.
Theorem (Gay Soliman FagesBioinformatics 26(18):i575i581, 2010) There exists a subgraph epimorphism from G to G0
if and only if there exists a graphical reduction from G to G0 by species/reactions deletions and mergings.
Subgraph isomorphisms(SISO) allow delete operations only Graph epimorphisms(GEPI) allow merge operations only
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Subgraph Epimorphisms
Definition
Asubgraph morphism(SMOR) fromG = (S,A) to G0= (S0,A0) is a functionµ:S0 −→S0, with S0 ⊆S such that
∀(x,y)∈A∩(S0×S0) (µ(x), µ(y))∈A0.
Asubgraph epimorphism (SEPI) is a surjective SMOR
∀x0 ∈S0 ∃x ∈S0 µ(x) =x0,
∀(x0,y0)∈A0 ∃(x,y)∈A µ(x) =x0 µ(y) =y0.
Theorem (Gay Soliman FagesBioinformatics 26(18):i575i581, 2010) There exists a subgraph epimorphism from G to G0
if and only if there exists a graphical reduction from G to G0 by species/reactions deletions and mergings.
Subgraph isomorphisms(SISO) allow delete operations only Graph epimorphisms(GEPI) allow merge operations only
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Subgraph Epimorphisms
Definition
Asubgraph morphism(SMOR) fromG = (S,A) to G0= (S0,A0) is a functionµ:S0 −→S0, with S0 ⊆S such that
∀(x,y)∈A∩(S0×S0) (µ(x), µ(y))∈A0.
Asubgraph epimorphism (SEPI) is a surjective SMOR
∀x0 ∈S0 ∃x ∈S0 µ(x) =x0,
∀(x0,y0)∈A0 ∃(x,y)∈A µ(x) =x0 µ(y) =y0.
Theorem (Gay Soliman FagesBioinformatics 26(18):i575i581, 2010) There exists a subgraph epimorphism from G to G0
if and only if there exists a graphical reduction from G to G0 by species/reactions deletions and mergings.
Subgraph isomorphisms(SISO) allow delete operations only Graph epimorphisms(GEPI) allow merge operations only
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Subgraph Epimorphisms
Definition
Asubgraph morphism(SMOR) fromG = (S,A) to G0= (S0,A0) is a functionµ:S0 −→S0, with S0 ⊆S such that
∀(x,y)∈A∩(S0×S0) (µ(x), µ(y))∈A0.
Asubgraph epimorphism (SEPI) is a surjective SMOR
∀x0 ∈S0 ∃x ∈S0 µ(x) =x0,
∀(x0,y0)∈A0 ∃(x,y)∈A µ(x) =x0 µ(y) =y0.
Theorem (Gay Soliman FagesBioinformatics 26(18):i575i581, 2010) There exists a subgraph epimorphism from G to G0
if and only if there exists a graphical reduction from G to G0 by species/reactions deletions and mergings.
Subgraph isomorphisms(SISO) allow delete operations only
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Model Reductions as Subgraph Epimorphisms
Example (Michaelis-Menten reduction) Subgraph epimorphism:
E →C S →A P →B c →r p →r d → ⊥ ES → ⊥
E c ES
d S
p P
A r
C
B Equivalent to the graphical reduction:
merge(c,p), delete(d), delete(ES)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
The SEPI Existence Problem
Input: two reaction graphs
Output: whether there exists a SEPI (i.e. a graphical model reduction) from the first graph to the second.
Theorem (Gay Martinez Fages Soliman SolnonDiscrete Applied Mathematics 162:214228, 2014)
The SEPI existence problem between two graphs is NP-complete.
Proof by reduction of the Set Covering Problem.
Practical significance of NP-completeness ?
For any algorithm there exists an infinite set of instances for which the computation time is exponential in their size. Says nothing about the existence of efficient algorithms for solving an infinite class of (practical) instances.
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
The SEPI Existence Problem
Input: two reaction graphs
Output: whether there exists a SEPI (i.e. a graphical model reduction) from the first graph to the second.
Theorem (Gay Martinez Fages Soliman SolnonDiscrete Applied Mathematics 162:214228, 2014)
The SEPI existence problem between two graphs is NP-complete.
Proof by reduction of the Set Covering Problem.
Practical significance of NP-completeness ?
For any algorithm there exists an infinite set of instances for which the computation time is exponential in their size.
Says nothing about the existence of efficient algorithms for solving an infinite class of (practical) instances.
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Constraint Satisfaction Problems
Aconstraint satisfaction problem(CSP for short) is a triple P = (X,D,C), where:
X is a set ofvariables
D is a family ofdomains (e.g.N) indexed by variables inX :
∀x ∈X,Dx is a finite set.
C is a set of constraints, eachc ∈C is defined by its arity ar(c)∈N, a tuple of variables~x(c)∈Xar(c) and a relation R(c)⊆Qar(c)
i=1 D~xi(c).
Anassignmentν :X −→Dx is a solutionof P when
∀c ∈C,(ν(x1), . . . , ν(xn))∈R(c), with~x(c) = (x1, . . . ,xn).
Example
For graph morphisms problems, associate one variable per vertex in the source graph, with the target graph vertices as domain.
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Constraint Satisfaction Problems
Aconstraint satisfaction problem(CSP for short) is a triple P = (X,D,C), where:
X is a set ofvariables
D is a family ofdomains (e.g.N) indexed by variables inX :
∀x ∈X,Dx is a finite set.
C is a set of constraints, eachc ∈C is defined by its arity ar(c)∈N, a tuple of variables~x(c)∈Xar(c) and a relation R(c)⊆Qar(c)
i=1 D~xi(c).
Anassignmentν :X −→Dx is a solutionof P when
∀c ∈C,(ν(x1), . . . , ν(xn))∈R(c), with~x(c) = (x1, . . . ,xn).
Example
For graph morphisms problems, associate one variable per vertex in the source graph, with the target graph vertices as domain.
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Domain Filtering Algorithms for Basic Constraints
A constraint is a reactive agent that reduces the domains of its variables upon domain reduction events.
Example (c =relation(~x,R))
where~x is a tuple of variables andR is a relation given in extension, is defined by:
ar(c) = arity of~x = arity ofR= n
~x(c) =~x R(c) =R ∩Qn
i=1Dxi,
Example (c =element(i,(f1, . . . ,fn),x) meaningfi =x) ar(c) =n+ 2
~x(c) = (i,f1, . . . ,fn,x)
R(c) ={(ι, ν1, . . . , νn, ξ)∈Di×(Qn
i=1Dfi)×Dx |νι =ξ}
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Domain Filtering Algorithms for Basic Constraints
A constraint is a reactive agent that reduces the domains of its variables upon domain reduction events.
Example (c =relation(~x,R))
where~x is a tuple of variables andR is a relation given in extension, is defined by:
ar(c) = arity of~x = arity ofR= n
~x(c) =~x R(c) =R ∩Qn
i=1Dxi,
Example (c =element(i,(f1, . . . ,fn),x) meaningfi =x) ar(c) =n+ 2
~x(c) = (i,f1, . . . ,fn,x)
R(c) ={(ι, ν1, . . . , νn, ξ)∈Di ×(Qn
i=1Dfi)×Dx |νι =ξ}
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
SEP Existence Problem as a CSP
Variablesfor the vertices ofG and G0 and for the edges of G0 X ={xv |v ∈V(G)} ] {yv0 |v0 ∈V(G0)} ] {ye0 |e0 ∈E(G0)}, D(xv) =V(G⊥0 ), D(yv0) ={1, . . . ,|V(G)|}, D(ye0) ={1, . . . ,|E(G)|}
Theconstraintsare :
{relation((xu,xv),E(G⊥0 ))|(u,v)∈E(G)}
∪{element(yv0,x(V(G)),v0)|v0 ∈V(G0)}
∪{element(y(u0,v0), π1(x(E(G))),u0)|(u0,v0)∈E(G0)}
∪{element(y(u0,v0), π2(x(E(G))),v0)|(u0,v0)∈E(G0)} wherex(V(G)) ={xv |v∈V(G)},
x(E(G)) ={(xu,xv) |(u,v)∈E(G)},
andπ1, π2 map the first and second projection functions on lists.
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
SEP Existence Problem as a CSP
Variablesfor the vertices ofG and G0 and for the edges of G0 X ={xv |v ∈V(G)} ] {yv0 |v0 ∈V(G0)} ] {ye0 |e0 ∈E(G0)}, D(xv) =V(G⊥0 ), D(yv0) ={1, . . . ,|V(G)|}, D(ye0) ={1, . . . ,|E(G)|}
Theconstraintsare :
{relation((xu,xv),E(G⊥0))|(u,v)∈E(G)}
∪{element(yv0,x(V(G)),v0)|v0 ∈V(G0)}
∪{element(y(u0,v0), π1(x(E(G))),u0)|(u0,v0)∈E(G0)}
∪{element(y(u0,v0), π2(x(E(G))),v0)|(u0,v0)∈E(G0)} wherex(V(G)) ={xv |v∈V(G)},
x(E(G)) ={(xu,xv) |(u,v)∈E(G)},
andπ1, π2 map the first and second projection functions on lists.
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Search Strategy
Enumerating source vertex variables is sufficient to enforce surjectivity and decide SEPI satisfiability.
In practice it is more efficient toenumerate the antecedent variables of the reduced graph
and finally instanciate the remaining source vertex variables to⊥. functionVariableValueSelectionantecedents(D)
if ∃v0 ∈V(G0)|yv0 ={d1,d2, . . .}then return(yv0,d1)
else if ∃(u0,v0)∈E(G0)|y(u0,v0) ={d1,d2, . . .} then return(y(u0,v0),d1)
else if ∃v ∈V(G)| |Dxv|>1∧Dxv 6={⊥} then return(xv,⊥)
end if end function
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Search Strategy
Enumerating source vertex variables is sufficient to enforce surjectivity and decide SEPI satisfiability.
In practice it is more efficient toenumerate the antecedent variables of the reduced graph
and finally instanciate the remaining source vertex variables to⊥.
functionVariableValueSelectionantecedents(D) if ∃v0 ∈V(G0)|yv0 ={d1,d2, . . .}then
return(yv0,d1)
else if ∃(u0,v0)∈E(G0)|y(u0,v0) ={d1,d2, . . .} then return(y(u0,v0),d1)
else if ∃v ∈V(G)| |Dxv|>1∧Dxv 6={⊥} then return(xv,⊥)
end if end function
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Evaluation on biomodels.net
Implemented in Biocham v3 (not yet in v4) in GNU-Prolog http://gprolog.inria.frand MiniSAT http://minisat.se
Search of a subgraph epimorphism between all pairs of models with a time out of 20mn
90% of comparisons take less than 5s
Computes model hierarchies where each node represents a model:
M −→M0 means a reduction fromM toM0 was found. M ←→M0 means M andM0 are isomorphic.
9% of false positive found between different model classes typically involving very small modelsrecognized as motifs in larger models (e.g. double phosphorylations)
1.2% of false positive after removal of small models
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Evaluation on biomodels.net
Implemented in Biocham v3 (not yet in v4) in GNU-Prolog http://gprolog.inria.frand MiniSAT http://minisat.se
Search of a subgraph epimorphism between all pairs of models with a time out of 20mn
90% of comparisons take less than 5s
Computes model hierarchies where each node represents a model:
M −→M0 means a reduction fromM toM0 was found.
M ←→M0 means M andM0 are isomorphic.
9% of false positive found between different model classes typically involving very small modelsrecognized as motifs in larger models (e.g. double phosphorylations)
1.2% of false positive after removal of small models
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Evaluation on biomodels.net
Implemented in Biocham v3 (not yet in v4) in GNU-Prolog http://gprolog.inria.frand MiniSAT http://minisat.se
Search of a subgraph epimorphism between all pairs of models with a time out of 20mn
90% of comparisons take less than 5s
Computes model hierarchies where each node represents a model:
M −→M0 means a reduction fromM toM0 was found.
M ←→M0 means M andM0 are isomorphic.
9% of false positive found between different model classes typically involving very small modelsrecognized as motifs in larger models (e.g. double phosphorylations)
1.2% of false positive after removal of small models
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
MAPK Hierarchy
009_Huan
010_Khol 011_Levc
027_Mark
029_Mark 031_Mark
026_Mark
028_Mark 030_Mark 049_Sasa
146_Hata
Models 009 (Huang 1996), 010 (Kholodenko 2000) and 011 (Levchenko 2000) arethree-level cascade models.
Models 026 to 031 (Markevitch 2004) areone-level.
Models 049 (Sasagawa 2005) is a larger model (216 reactions), some computations timed out.
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Circadian Clock Models Hierarchy
021_Lelo
170_Weim
022_Ueda
034_Smol
055_Lock 073_Lelo
078_Lelo 074_Lelo
083_Lelo
089_Lock 171_Lelo
Models 073, 078 isomorphic (different parameter values).
Models 074, 083 isomorphic, refinement of 073 078 with RevErbα False negative: models 021, 171 have the same ODEs but with different reaction encodings(variable eliminated by invariants)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Calcium Oscillation Models Hierarchy
039_Marh
098_Gold
115_Somo
117_Dupo
166_Zhu 043_Borg
044_Borg
045_Borg
058_Bind 122_Fish
145_Wang
Models 098, 115, 117 are very small two-species oscillators.
Model 122 (Fisher et al. 2006) NFAT, NFκB and side calcium
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Cell Cycle Models Hierarchy
007_Nova
008_Gard 168_Obey
056_Chen
169_Agud 109_Habe 111_Nova 196_Sriv
144_Calz
Not satisfactory: these ODE models have been transcribed in SBML withill-formed reactions (missing reactants, same ODEs) Species eliminated byconservation lawsare encoded by terms in the kinetics and invisible in the reaction
Events(for cell division) are invisible in the reaction hypergraph.
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Problem of ODE Models badly Transcribed in SBML
E.g. Biomodels 8 originiates from an ODE model of a simple mitotic oscillator [Goldbeter 91]
It wasquickly transcribed in SBML with synthesis and degradation reactions and keeping terms for molecules eliminated by invariants M0 = 1−M X0= 1−X.
Can we infer the right structure of the SBML reactions from the ODE? infer hidden molecules andwell-formed reactions...
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Problem of ODE Models badly Transcribed in SBML
E.g. Biomodels 8 originiates from an ODE model of a simple mitotic oscillator [Goldbeter 91]
It wasquickly transcribed in SBML with synthesis and degradation reactions and keeping terms for molecules eliminated by invariants M0 = 1−M X0= 1−X.
Can we infer the right structure of the SBML reactions from the ODE? infer hidden molecules andwell-formedreactions...
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Infering Reaction Systems from ODEs
dAi/dt =fi(A) is the ODE semantics offi for => Ai
Definition
A reaction with inhibitors m,f for r / m => p, over molecular species{x1, . . . ,xs} is well-formed if the following conditions hold:
1 f(x1, . . . ,xs) is partially differentiable, non-negative on Rs+;
2 xi ∈r if and only if∂f/∂xi(~x)>0 for some value~x ∈Rs+;
3 xi ∈m if and only if∂f/∂xi(~x)<0 for some value~x ∈Rs+. A molecular species can be both areactant and aninhibitor. Example
To ˙x =−k one can associate the well-formed reaction system l*x for x => 2*x,k+l*x for x =>
This is the result computed in BIOCHAM byload ode. Note thatk+l∗x6= 0 when x = 0
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Infering Reaction Systems from ODEs
dAi/dt =fi(A) is the ODE semantics offi for => Ai Definition
A reaction with inhibitors m,f for r / m => p, over molecular species{x1, . . . ,xs} is well-formed if the following conditions hold:
1 f(x1, . . . ,xs) is partially differentiable, non-negative on Rs+;
2 xi ∈r if and only if∂f/∂xi(~x)>0 for some value~x ∈Rs+;
3 xi ∈m if and only if∂f/∂xi(~x)<0 for some value~x ∈Rs+.
A molecular species can be both areactant and aninhibitor. Example
To ˙x =−k one can associate the well-formed reaction system l*x for x => 2*x,k+l*x for x =>
This is the result computed in BIOCHAM byload ode. Note thatk+l∗x6= 0 when x = 0
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Infering Reaction Systems from ODEs
dAi/dt =fi(A) is the ODE semantics offi for => Ai Definition
A reaction with inhibitors m,f for r / m => p, over molecular species{x1, . . . ,xs} is well-formed if the following conditions hold:
1 f(x1, . . . ,xs) is partially differentiable, non-negative on Rs+;
2 xi ∈r if and only if∂f/∂xi(~x)>0 for some value~x ∈Rs+;
3 xi ∈m if and only if∂f/∂xi(~x)<0 for some value~x ∈Rs+. A molecular species can be both areactantand aninhibitor.
Example
To ˙x =−k one can associate the well-formed reaction system
l*x for x => 2*x,k+l*x for x =>
This is the result computed in BIOCHAM byload ode. Note thatk+l∗x6= 0 when x = 0
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Infering Reaction Systems from ODEs
dAi/dt =fi(A) is the ODE semantics offi for => Ai Definition
A reaction with inhibitors m,f for r / m => p, over molecular species{x1, . . . ,xs} is well-formed if the following conditions hold:
1 f(x1, . . . ,xs) is partially differentiable, non-negative on Rs+;
2 xi ∈r if and only if∂f/∂xi(~x)>0 for some value~x ∈Rs+;
3 xi ∈m if and only if∂f/∂xi(~x)<0 for some value~x ∈Rs+. A molecular species can be both areactantand aninhibitor.
Example
To ˙x =−k one can associate the well-formed reaction system l*x for x => 2*x,k+l*x for x =>
This is the result computed in BIOCHAM byload ode.
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Strict Kinetics and Positive Systems
A reactionf for r / m => p isstrict if its kinetics
f(x1, . . . ,xs) = 0 wheneverxj = 0 for anyxj such that r(xj)>0.
A dynamical system overRk is called positive ifRk+ is an invariant set for the system, i.e.,∀x0≥0,t ≥0x(t,x0)≥0
Proposition (Positivity)
Well-formed and strict reactions define positive ODEs Proof: We have ˙xj=Pn
i=1(pi(xj)−ri(xj))×fi and since the system is well-formed, thefi are all non-negative. The only negative terms thus haveri(xj)>0 and from the strictness condition this entails thatfi= 0 whenxj = 0. Hence ˙xj ≥0 wheneverxj = 0 since it is a sum of
non-negative terms. Thereforexj cannot become negative when its initial value is non-negative, and since this holds for allj, the system is positive.
˙
x=−k is not the ODE semantics of strict well-formed reactions
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Strict Kinetics and Positive Systems
A reactionf for r / m => p isstrict if its kinetics
f(x1, . . . ,xs) = 0 wheneverxj = 0 for anyxj such that r(xj)>0.
A dynamical system overRk is called positive ifRk+ is an invariant set for the system, i.e.,∀x0 ≥0,t≥0x(t,x0)≥0
Proposition (Positivity)
Well-formed and strict reactions define positive ODEs
Proof: We have ˙xj=Pn
i=1(pi(xj)−ri(xj))×fi and since the system is well-formed, thefi are all non-negative. The only negative terms thus haveri(xj)>0 and from the strictness condition this entails thatfi= 0 whenxj = 0. Hence ˙xj ≥0 wheneverxj = 0 since it is a sum of
non-negative terms. Thereforexj cannot become negative when its initial value is non-negative, and since this holds for allj, the system is positive.
˙
x=−k is not the ODE semantics of strict well-formed reactions
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Strict Kinetics and Positive Systems
A reactionf for r / m => p isstrict if its kinetics
f(x1, . . . ,xs) = 0 wheneverxj = 0 for anyxj such that r(xj)>0.
A dynamical system overRk is called positive ifRk+ is an invariant set for the system, i.e.,∀x0 ≥0,t≥0x(t,x0)≥0
Proposition (Positivity)
Well-formed and strict reactions define positive ODEs Proof: We have ˙xj=Pn
i=1(pi(xj)−ri(xj))×fi and since the system is well-formed, thefi are all non-negative. The only negative terms thus haveri(xj)>0 and from the strictness condition this entails thatfi= 0 whenxj = 0. Hence ˙xj≥0 wheneverxj = 0 since it is a sum of
non-negative terms. Thereforexj cannot become negative when its initial value is non-negative, and since this holds for allj, the system is positive.
˙
x=−k is not the ODE semantics of strict well-formed reactions
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Strict Kinetics and Positive Systems
A reactionf for r / m => p isstrict if its kinetics
f(x1, . . . ,xs) = 0 wheneverxj = 0 for anyxj such that r(xj)>0.
A dynamical system overRk is called positive ifRk+ is an invariant set for the system, i.e.,∀x0 ≥0,t≥0x(t,x0)≥0
Proposition (Positivity)
Well-formed and strict reactions define positive ODEs Proof: We have ˙xj=Pn
i=1(pi(xj)−ri(xj))×fi and since the system is well-formed, thefi are all non-negative. The only negative terms thus haveri(xj)>0 and from the strictness condition this entails thatfi= 0 whenxj = 0. Hence ˙xj≥0 wheneverxj = 0 since it is a sum of
non-negative terms. Thereforexj cannot become negative when its initial value is non-negative, and since this holds for allj, the system is positive.
˙
x=−k is not the ODE semantics of strict well-formed reactions
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Reaction Inference Algorithm
input: ODE systemOover variables for molecular concentrations, partial has pos valtest
1 rewriteOinto additive normal form 2 compute the setT of all terms appearing inO 3 letR:=∅
4 for each non-decomposable termtinT, 1 letr:=∅ ,p:=∅ ,m:=∅
2 for each variablex wheret occurs with integer coefficient c in
˙ x inO,
1 ifc<0 thenr(x) :=−c,
2 ifc>0 thenp(x) :=c,
3 for each variablex such thatr(x) = 0 and partial has pos val(t,x),
1 r(x) := 1,
2 p(x) :=p(x) + 1,
4 for each variablex such thatpartial has pos val(−t,x),
1 m(x) := 1,
5 R:=R∪ {r /m−→t p},
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Inference of Molecules Eliminated by Invariants
input: ODE systemOover variables{x1, . . . ,xs},
1 iteratively replace inOany expression of the form−x+ybyy−x,
2 for each expression of the formk−x−yinOwherekis a numerical constant or a parameter, andxand yare variables,
1 introduce a new variablez with time derivative ˙z =−x˙−y˙, and functional dependency equationz =k−x−y,
2 substitute any occurrence ofk−x−y in O byz,
3 substitute any occurrence ofk+v−x−y inO for any expressionv, by v+z,
4 substitute any occurrence ofk−x+w−y inO for any w, by v+z,
3 for each expressionk−xappearing inOwherekis a constant or a parameter andxa variable, 1 introduce a new variablez with time derivative ˙z =−x˙ and
functional dependency equationz =k−x,
2 substitute any occurrence ofk−x inO byz,
3 substitute any occurrence ofk+v−x in O for any expression v, by z+v,
output: ODE systemOover variables{x1, . . . ,xs}and hidden molecule variables{z1, . . . ,zk}, together with
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Conclusion
To a large extend, model reductions can be detected in model repositories by checking the existence of SEPI between the reaction hypergraphs
The SEPI existence problem between two graphs is
NP-complete but constraint programming and SAT solvers apply with few timeouts
Automatic creation of model hierarchies (metamodels) pauseitem Works only for well-written SBML models (well-formed reaction systems)
Automatic inference of well-formed reaction systems from ODEs (Biocham load odecommand, used to import MatLab models in SBML)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Conclusion
To a large extend, model reductions can be detected in model repositories by checking the existence of SEPI between the reaction hypergraphs
The SEPI existence problem between two graphs is
NP-complete but constraint programming and SAT solvers apply with few timeouts
Automatic creation of model hierarchies (metamodels) pauseitem Works only for well-written SBML models (well-formed reaction systems)
Automatic inference of well-formed reaction systems from ODEs (Biocham load odecommand, used to import MatLab models in SBML)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Conclusion
To a large extend, model reductions can be detected in model repositories by checking the existence of SEPI between the reaction hypergraphs
The SEPI existence problem between two graphs is
NP-complete but constraint programming and SAT solvers apply with few timeouts
Automatic creation of model hierarchies (metamodels) pauseitem Works only for well-written SBML models (well-formed reaction systems)
Automatic inference of well-formed reaction systems from ODEs (Biocham load odecommand, used to import MatLab models in SBML)
Model reductions Graph Reductions Graph Morphisms Algorithms Evaluation in BioModels Well-formed Reactions
Conclusion
To a large extend, model reductions can be detected in model repositories by checking the existence of SEPI between the reaction hypergraphs
The SEPI existence problem between two graphs is
NP-complete but constraint programming and SAT solvers apply with few timeouts
Automatic creation of model hierarchies (metamodels) pauseitem Works only for well-written SBML models (well-formed reaction systems)
Automatic inference of well-formed reaction systems from ODEs (Biochamload ode command, used to import MatLab models in SBML)