Satisfying Flux Balance and Mass-Action Kinetics in a Network of Biochemical Reactions

Transcription

1 Satisfying Flux Balance and Mass-Action Kinetics in a Network of Biochemical Reactions Speakers Senior Personnel Collaborators Michael Saunders and Ines Thiele Ronan Fleming, Bernhard Palsson, Yinyu Ye Santiago Akle, Onkar Dalal, Joshua Lerman, Yuekai Sun, Nicole Taheri ICME Systems Biology Research Group Center for Systems Biology Stanford University University of California, San Diego University of Iceland Genomic Sciences Annual PI Meeting Hyatt Regency Crystal City, VA April 1013, 2011 Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 1/19

2 Abstract Computational modeling of biochemical reaction networks will become increasingly dependent on large-scale numerical optimization as the network dimensionality continues to rise. Linear optimization, the mainstay for ux balance analysis, is trivial for well scaled metabolic models, but nontrivial for integrated metabolic and macromolecular synthetic networks because of the size of the stoichiometric matrix and the wide temporal scales involved. We describe the rst such reconstruction for E. coli, which accounts for 43% of all its genes, and a similar reconstruction for T. maritima as well as numerical challenges associated with ux balance analysis. Beyond linear optimization, we have established that a sequence of parametric convex optimizations can be used to model the inherently nonlinear kinetic relationship between reaction ux and metabolite concentration. Convex optimization preserves many of the attractive numerical features of linear optimization algorithms. For example, given a model of a biochemical network, one can establish a priori whether a solution will exist. The non-existence of a solution is an indication of a malformed model arising from incorrect stoichiometry, reaction directionality, etc. Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 2/19

3 Abstract, contd With a sequence of convex optimization problems, establishing the existence of a xed point for such a sequence is equally important for indicating when a model is malformed. We summarize recent results establishing necessary conditions for existence of a ux and concentration at such a xed point. The approach also leads to a numerically tractable algorithm for simultaneous satisfaction of ux balance and mass-action kinetics for a well scaled stoichiometric model, subject to certain conditions. The overall objective of this work is to provide the systems biology community with a new wave of computationally ecient algorithms for physico-chemically realistic modeling of genome-scale biochemical networks. We highlight open source algorithms already released. Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 3/19

4 1 Flux Balance Analysis LP problem 2 Entropy problem 3 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems 4 LUSOL Other solvers Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 4/19

5 Flux Balance Analysis LP problem Leads to linear programming problems involving the stoichiometric matrix S (Palsson 2006): minimize v f, v r, v e d T v e subject to Sv f Sv r + S e v e = 0 v f, v r 0, l v e u Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 5/19

6 Flux Balance Analysis LP problem Leads to linear programming problems involving the stoichiometric matrix S (Palsson 2006): minimize v f, v r, v e d T v e subject to Sv f Sv r + S e v e = 0 v f, v r 0, l v e u Treat as a challenging large-scale linear program: minimize x d T x subject to Ax = b, l x u Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 5/19

7 LP problem Badly scaled stoichiometric constraints When modeling biological systems, we often encounter reactions with large coecients: A B C + D We replace these reactions with a sequence of reactions A ˆB 2 C + D 100 ˆB 1 ˆB 2 100B ˆB 1 Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 6/19

8 Badly scaled coupling constraints LP problem Recently, reaction coupling has been used to represent dependencies between synthesis and utilization [Thiele et al., Biophys. J., 98(10), 2072 (2010)]: α v 1 v 2 β Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 7/19

9 Badly scaled coupling constraints LP problem Recently, reaction coupling has been used to represent dependencies between synthesis and utilization [Thiele et al., Biophys. J., 98(10), 2072 (2010)]: These are linear constraints: α v 1 v 2 β αv 2 v 1 0 v 1 βv 2 0 Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 7/19

10 Badly scaled coupling constraints LP problem Recently, reaction coupling has been used to represent dependencies between synthesis and utilization [Thiele et al., Biophys. J., 98(10), 2072 (2010)]: These are linear constraints: α v 1 v 2 β αv 2 v 1 0 v 1 βv 2 0 We replace v v 2 0 by a sequence of constraints: v 1 100s 1 0 s 1 100s 2 0 s 2 100v 2 0 Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 7/19

11 Example data Flux Balance Analysis LP problem Ax = b Organism Problem Dimensions T. maritima ME_ E. coli ME_ E. coli ME_ CPLEX simplex and barrier LP solvers applied with Feasibility and Optimality tolerances 10 9 Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 8/19

12 Experimental results LP problem CPLEX simplex solver original with rebuild Problem b Ax 1 Time b Ax 1 Time TM EC EC Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 9/19

13 Experimental results LP problem CPLEX simplex solver original with rebuild Problem b Ax 1 Time b Ax 1 Time TM EC EC CPLEX barrier solver original with rebuild Problem b Ax 1 Time b Ax 1 Time TM EC EC Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 9/19

14 LP problem Implications for Flux Balance Analysis Our procedure addresses the numerical diculties associated with badly scaled linear systems that arise in systems biology Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 10/19

15 LP problem Implications for Flux Balance Analysis Our procedure addresses the numerical diculties associated with badly scaled linear systems that arise in systems biology Users can more accurately model biological systems by enforcing ill-scaled constraints Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 10/19

16 LP problem Implications for Flux Balance Analysis Our procedure addresses the numerical diculties associated with badly scaled linear systems that arise in systems biology Users can more accurately model biological systems by enforcing ill-scaled constraints Remember to switch o the CPLEX Presolve Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 10/19

17 LP problem Implications for Flux Balance Analysis Our procedure addresses the numerical diculties associated with badly scaled linear systems that arise in systems biology Users can more accurately model biological systems by enforcing ill-scaled constraints Remember to switch o the CPLEX Presolve A further card up sleeve: 128-bit f90 version of SQOPT Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 10/19

18 LP problem Implications for Flux Balance Analysis Our procedure addresses the numerical diculties associated with badly scaled linear systems that arise in systems biology Users can more accurately model biological systems by enforcing ill-scaled constraints Remember to switch o the CPLEX Presolve A further card up sleeve: 128-bit f90 version of SQOPT Avoiding large numbers in S is essential for rank estimation (e.g. to evaluate conserved moieties correctly) Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 10/19

19 Entropy problem Entropy problem with parameter p minimize v f, v r>0 v T f (log v f + p e) + v T r (log v r + p e) subject to Sv f Sv r = b : y Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 11/19

20 Entropy problem Entropy problem with parameter p minimize v f, v r>0 v T f (log v f + p e) + v T r (log v r + p e) subject to Sv f Sv r = b : y Optimality condition S T y log(v r./ v f ) Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 11/19

21 Entropy problem Entropy problem with parameter p minimize v f, v r>0 v T f (log v f + p e) + v T r (log v r + p e) subject to Sv f Sv r = b : y Optimality condition S T y log(v r./ v f ) Optimal v f, v r are thermodynamically feasible First numerically scalable method (polynomial complexity) [Fleming et al., A variational principle for computing nonequilibrium uxes and potentials in genome-scale biochemical networks, 2011] Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 11/19

22 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems Homogeneous network (stepping stone to b 0) [ ] [ ] v S S f = 0 S R F v r Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 12/19

23 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems Homogeneous network (stepping stone to b 0) [ ] [ ] v S S f = 0 S R F v r (R F )v f = (R F )v r v f = diag (k f ) v r =... i cf i1 i. i cf in i Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 12/19

24 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems Homogeneous network (stepping stone to b 0) [ ] [ ] v S S f = 0 S R F v r (R F )v f = (R F )v r v f = diag (k f ) v r =... F v r + Rv f = F v f + Rv r v(c) = i cf i1 i. i cf in i i cy i1 i. i cy in i Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 12/19

25 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems Homogeneous network (stepping stone to b 0) [ ] [ ] v S S f = 0 S R F v r (R F )v f = (R F )v r v f = diag (k f ) v r =... F v r + Rv f = F v f + Rv r v(c) = i cf i1 i. i cf in i i cy i1 i. i cy in i Y A T v(c) = Y Dv(c) log v(c) = Y T log c Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 12/19

26 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems We want Y Dv = Y A T v and Y T log c = log v Entropy problem with parameter µ minimize v>0 v T D log v e T Dv subject to Y Dv = µ : y Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 13/19

27 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems We want Y Dv = Y A T v and Y T log c = log v Entropy problem with parameter µ minimize v>0 v T D log v e T Dv subject to Y Dv = µ : y Optimality condition DY T y(µ) = D log v(µ) Y T y(µ) = log v(µ) Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 13/19

28 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems We want Y Dv = Y A T v and Y T log c = log v Entropy problem with parameter µ minimize v>0 v T D log v e T Dv subject to Y Dv = µ : y Optimality condition DY T y(µ) = D log v(µ) Y T y(µ) = log v(µ) We would like the xed point condition µ = Y A T v(µ) Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 13/19

29 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems Fixed point iteration v v 0 µ Y A T v while Y Dv Y A T v > τ do Solve Entropy subproblem min v T D log v e T Dv st Y Dv = µ end µ (1 θ)µ + θy A T v Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 14/19

30 Fixed point iteration, θ = 0.5 Reformulation of Sv f Sv r = 0 Parameterized Entropy problem Sequence of Entropy problems Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 15/19

31 LUSOL: A sparse LU package LUSOL Other solvers Factorize m n sparse matrix S = LDU with permutations to preserve stability and sparsity L and U triangular L ii = 1, D diagonal, U ii = 1 Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 16/19

32 LUSOL: A sparse LU package LUSOL Other solvers Factorize m n sparse matrix S = LDU with permutations to preserve stability and sparsity L and U triangular L ii = 1, D diagonal, U ii = 1 Uses: Find rank(s) < m < n Remove redundant rows Form nullspace basis Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 16/19

33 LUSOL: A sparse LU package LUSOL Other solvers Factorize m n sparse matrix S = LDU with permutations to preserve stability and sparsity L and U triangular L ii = 1, D diagonal, U ii = 1 Uses: Find rank(s) < m < n Remove redundant rows Form nullspace basis Threshold partial pivoting keeps L well-conditioned L ij Ltol (5.0 or 2.0 say) Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 16/19

34 LUSOL: A sparse LU package LUSOL Other solvers Factorize m n sparse matrix S = LDU with permutations to preserve stability and sparsity L and U triangular L ii = 1, D diagonal, U ii = 1 Uses: Find rank(s) < m < n Remove redundant rows Form nullspace basis Threshold partial pivoting keeps L well-conditioned L ij Ltol (5.0 or 2.0 say) Threshold rook pivoting controls L and U L ij and U ij Ltol (2.0 or 1.5 say) Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 16/19

35 LUSOL Other solvers Stoichiometric matrix (rank-decient) 2 S = Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 17/19

36 LUSOL Other solvers Stoichiometric matrix (rank-decient) 2 S = SVD S = UDV T U, V orthogonal D = ɛ 1 Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 17/19

37 LUSOL Other solvers Stoichiometric matrix (rank-decient) 2 S = SVD S = UDV T U, V orthogonal D = ɛ 1 LUSOL with threshold rook pivoting Ltol= S = LDU = ɛ x x x Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 17/19

38 LUSOL Other solvers LUSOL with rook pivoting and reducing Ltol Sorted elements of D: SVD ɛ Ltol ɛ ɛ ɛ Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 18/19

39 Available software LUSOL Other solvers LUSOL: Fortran 77, C, f90 coming New Matlab mex interface (Nick Henderson, ICME, Stanford) Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 19/19

40 Available software LUSOL Other solvers LUSOL: Fortran 77, C, f90 coming New Matlab mex interface (Nick Henderson, ICME, Stanford) Nullspace functions: Compute products p = Zv or q = Z T y, where S T Z = 0 Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 19/19

41 Available software LUSOL Other solvers LUSOL: Fortran 77, C, f90 coming New Matlab mex interface (Nick Henderson, ICME, Stanford) Nullspace functions: Compute products p = Zv or q = Z T y, where S T Z = 0 PDCO: Matlab Barrier method for LP and Entropy problems Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 19/19

42 Available software LUSOL Other solvers LUSOL: Fortran 77, C, f90 coming New Matlab mex interface (Nick Henderson, ICME, Stanford) Nullspace functions: Compute products p = Zv or q = Z T y, where S T Z = 0 PDCO: Matlab Barrier method for LP and Entropy problems MINRES, SYMMLQ, LSQR, LSMR: f77, f90, Matlab Iterative solvers for Ax = b, min Ax b Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 19/19

43 Available software LUSOL Other solvers LUSOL: Fortran 77, C, f90 coming New Matlab mex interface (Nick Henderson, ICME, Stanford) Nullspace functions: Compute products p = Zv or q = Z T y, where S T Z = 0 PDCO: Matlab Barrier method for LP and Entropy problems MINRES, SYMMLQ, LSQR, LSMR: f77, f90, Matlab Iterative solvers for Ax = b, min Ax b SQOPT, SNOPT: f77, f90 coming General-purpose solvers for LP, QP, NLP Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 19/19

44 Available software LUSOL Other solvers LUSOL: Fortran 77, C, f90 coming New Matlab mex interface (Nick Henderson, ICME, Stanford) Nullspace functions: Compute products p = Zv or q = Z T y, where S T Z = 0 PDCO: Matlab Barrier method for LP and Entropy problems MINRES, SYMMLQ, LSQR, LSMR: f77, f90, Matlab Iterative solvers for Ax = b, min Ax b SQOPT, SNOPT: f77, f90 coming General-purpose solvers for LP, QP, NLP Systems Optimization Laboratory Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 19/19

45 Available software LUSOL Other solvers LUSOL: Fortran 77, C, f90 coming New Matlab mex interface (Nick Henderson, ICME, Stanford) Nullspace functions: Compute products p = Zv or q = Z T y, where S T Z = 0 PDCO: Matlab Barrier method for LP and Entropy problems MINRES, SYMMLQ, LSQR, LSMR: f77, f90, Matlab Iterative solvers for Ax = b, min Ax b SQOPT, SNOPT: f77, f90 coming General-purpose solvers for LP, QP, NLP Systems Optimization Laboratory and more (COBRA Toolbox and extensions) Stanford/UCSD/Iceland Satisfying Flux Balance and Mass-Action Kinetics 19/19