A possibilistic framework for constraintbased metabolic flux analysis
 Francisco Llaneras^{1}Email author,
 Antonio Sala^{1} and
 Jesús Picó^{1}
DOI: 10.1186/17520509379
© Llaneras et al; licensee BioMed Central Ltd. 2009
Received: 24 March 2009
Accepted: 31 July 2009
Published: 31 July 2009
Abstract
Background
Constraintbased models allow the calculation of the metabolic flux states that can be exhibited by cells, standing out as a powerful analytical tool, but they do not determine which of these are likely to be existing under given circumstances. Typical methods to perform these predictions are (a) flux balance analysis, which is based on the assumption that cell behaviour is optimal, and (b) metabolic flux analysis, which combines the model with experimental measurements.
Results
Herein we discuss a possibilistic framework to perform metabolic flux estimations using a constraintbased model and a set of measurements. The methodology is able to handle inconsistencies, by considering sensors errors and model imprecision, to provide rich and reliable flux estimations. The methodology can be cast as linear programming problems, able to handle thousands of variables with efficiency, so it is suitable to deal with largescale networks. Moreover, the possibilistic estimation does not attempt necessarily to predict the actual fluxes with precision, but rather to exploit the available data – even if those are scarce – to distinguish possible from impossible flux states in a gradual way.
Conclusion
We introduce a possibilistic framework for the estimation of metabolic fluxes, which is shown to be flexible, reliable, usable in scenarios lacking data and computationally efficient.
Background
Systems biology states that, in order to quantitatively understand and predict the cell behaviour, its constitutive components and their interactions must be studied as a whole system [1, 2]. Metabolic networks are a paradigmatic example of this aim because, even incomplete as they may be, they are the best characterized cellular networks [3]. In recent times, the information embedded in metabolic networks is being used to assemble constraintbased models under the pseudo steadystate assumption, thus not requiring the knowledge of kinetic parameters, which are still rarely known [3, 4]. Constraintbased models allow the calculation of the possible metabolic states or "behaviours" that can be exhibited by the cell; however, they do not predict which of these are likely under given circumstances. One approach to perform these predictions is flux balance analysis (FBA), which is based on the assumption that cell behaviour has evolved to be optimal in a certain sense [5, 6]. It has been shown that FBA is able to predict the actual fluxes [7–9], but this requires to identify which are the relevant objectives for different conditions [7, 10]. As an alternative, one could perform a metabolic flux analysis (MFA) which, generally speaking, is the exercise of estimating the fluxes shown by cells by combination of a constraintbased model and the set of available experimental measurements.
In order to estimate the intracellular fluxes, traditional metabolic flux analysis (TMFA) employs only measurements of uptake and production rates (i.e. influxes into and outfluxes from cells) that are stoichiometrically balanced [11]. This purely stoichiometric approach has some limitations, but most of them can be overcome with simple extensions, as it will be shown below.
One typical difficulty to be tackled by MFA is that the available measurements may be insufficient to estimate the intracellular fluxes, particularly in largescale networks, because there may be different flux distributions compatible with the available measurements. To face this situation, intracellular information obtained from stable isotope tracer experiments has been incorporated in many studies (13CMFA) [12–14]. Yet, data from isotope tracer experiments will not be considered in this work. Instead, we follow a constraintbased modeling approach, in the sense that we do not attempt necessarily to predict the actual fluxes with precision, but rather to distinguish "most possible" from "impossible" flux states, based on a suitable definition of "possibility", a constraintbased model and the available measurements, which in most cases do not include isotopic data.
Another option to face a lack of measurements is the use of some rational hypotheses to chose one flux distribution among those that are compatible with the measurements. For instance, Nookaew et al. have proposed to estimate the intracellular fluxes based on the assumption that cells are likely to use as many pathways as possible to maintain robustness and redundancy [15]. Related hypotheses have been formulated using the concept of elementary modes [16, 17]. The assumption of optimal cell behavior typically used in FBA could be also used (e.g. [7]). It will be shown that the methodology we propose is able to detect these flux distribution that are equally possible (or similarly possible), but for the sake of simple exposition we will not use any hypothesis herein. However, the possibilistic framework might be extended to incorporate hypotheses, as discussed in the conclusions section.
In this context, the paper discusses the use of a possibilistic framework for metabolic flux analysis.
Uncertainty, lack of measurements and model imprecision will be handled introducing the notion of "degree of possibility". Then, an efficient optimizationbased approach will be employed to query the most possible fluxes and their possibility distributions. The methodology is based on a reinterpretation of the consistent causal reasoning paradigm [18] as an equivalent problem of feasibility subject to equality and inequality constraints; preferences under uncertain knowledge are incorporated by transforming the feasibility problem into a linear optimisation one, which may be interpreted in possibilistic terms. The optimisation approach to logic reasoning has been previously explored by the research group to which the authors belong in [19–21], and this paper applies it to MFA.
The main features of the possibilistic framework introduced in the paper are the following: (i) it is based on a constraintbased model and not only on stoichiometric balances, (ii) it considers measurements uncertainty in a flexible way (e.g. nonsymmetric error or a band of uncertainty due to systemic error) and (iii) even model imprecision, (iv) it provides possibility distributions (and intervals) which are more informative than pointwise estimations when multiple flux values might be reasonably possible, (v) it is reliable even if only a few fluxes are measurable, (vi) it has the ability to detect, and handle, inconsistencies between measurements and model, and furthermore (vii) with high computational efficiency.
The structure of the paper is as follows: preliminaries on possibility, optimization and metabolic flux analysis are first addressed. Then, the basics of Possibilistic MFA and some refinements are discussed; the framework is illustrated with simple examples and a wellknow model of C. glutamicum. The paper is closed with a summary and a discussion on future work.
Preliminaries: possibility and optimization
In an abstract ideal situation, many estimation problems in science and engineering can be cast as estimating some decision variables δ given the known values of a set of other ones m (possibly, measurements) and a model expressed as a set of equality and inequality constraints (involving decision variables, measurements and some model parameters). Then, the valid estimations will be the feasible solutions of a constraint satisfaction problem [22, 23].
However, in many practical cases, the measurements are imprecise and the model parameters and constraints are also not accurate, so real data violates them. This is the reason why most reallife models should include uncertainty. The most basic representation of uncertainty would be giving interval values to measurements and model parameters. Refinements of the uncertainty representation give rise to probabilistic [23–25] and possibilistic [26–28] frameworks.
Probabilistic frameworks have an underlying interpretation in terms of the frequency in which some flux conditions appear; on the other hand, possibilistic frameworks measure the degree of compliance (consistency) of some decision variables with some (soft) modeling constraints. In this sense, the basic assumptions of both paradigms of inference under uncertainty are different.
In the following subsections the possibilistic framework will be described. Afterwards, this section of preliminaries will be closed discussing the relationship between probability and possibility and justifying the use the possibilistic framework.
Soft constraint satisfaction problems: a possibilistic approach
As explained above, the possibilistic framework is the chosen representation for the problem under study, following the ideas in [29], where possibilistic constraint satisfaction problems (CSP) are presented. There, the authors introduce constraints which are satisfied to a degree, transforming the feasibility/infeasibility of a potential solution into a gradual notion: given a CSP with multiple solutions δ ∈ Δ (where Δ denotes the search space over which feasible values for the decision variables will be searched), a function π : Δ → [0, 1] was suggested in order to represent preference or priority as a "consistency degree". The meaning of π(δ) = 1 would indicate that δ is in full agreement with the model and measurement constraints; the meaning of π(δ) = 0 indicates that δ is in "absolute, total contradiction" with the problem constraints, and never should be considered a feasible value. Intermediate values would denote values of decision variables which "somehow mildly" violate the problem constraints but could be considered "partially possible" from the "practical" knowledge of the "expert" modeller who defined π. The higher the value of π(δ), the higher the accordance with the problem constraints should be (subjectively interpreted as a higher "possibility" of the decision variable choice δ). Given the here outlined subjective meaning of π, it is denoted in literature as possibility distribution. The possibilistic calculus [27, 29] refers then, to computations with possibility distributions from a series of axioms. Basic ideas on it will be outlined below in this section. A simple example now illustrates the basic idea.
Example
Consider a flux balance {f_{1} = f_{2}}, stating equality between two flows, f_{1} and f_{2}, supposedly measured in a biological or chemical reaction. The measurements m_{ a }= (5, 7) and m_{ b }= (5, 5.1) are infeasible, whereas m_{ c }= (5, 5) is feasible. However, it seems clear that the subjective "possibility" of m_{ b }is higher than that of m_{ a }; m_{ b }can be thought to be quite reasonable in practice due to measurement errors.
The idea can be easily formalised for further computations by defining a possibility distribution, for instance, . In this way, potential solutions can be ranked: π(m_{ a }) = 0.018, π (m_{ b }) = 0.99 and π(m_{ c }) = 1. The search space in which to define the possibility, Δ, could be defined as, say, Δ = {(δ_{1}, δ_{2})0 ≤ δ_{ i }≤ 10}.
Usually, the function π(δ) is built by "conjunction" of possibility functions of individual relations π_{ i }(δ_{ i }) (expressing userdefined preference or priority on each individual constraint, in many cases in a problemdependent way). Such conjunction will be latter discussed in this section. The best CSP solutions are defined to be those which satisfy the global problem to the maximal degree.
In this way, once the user has defined such function expressing how a particular combination of system variables is "consistent" with its model, the basic idea on possibilistic calculus is, given a subset of the system variables (assumed as known or measured), estimate the "most possible" values of all the remaining variables via an optimization problem. The close relationship between possibilistic calculus and optimisation is discussed in the subsection below.
Possibility theory
The basic building block of possibility theory is a userdefined possibility distribution π : Δ → [0, 1]. This defines the possibility of each "point" δ in Δ. A consistent problem formulation is defined to be the one in which there exists at least one point with possibility equal to one.
The second building block are events, formally defined as subsets of Δ, in order to address problems such as, in the above example, determining the possibility of event A = {(f_{1}, f_{2}) ∈ Δ  0 ≤ f_{1} ≤ 3, 4 ≤ f_{2} ≤ 10}.
Possibility calculus as optimization
and, obviously, given two events A and B, A ⊂ B entails π(A) ≤ π(B).
Hence, possibility computations are optimisation problems (Cf. with probability computations, which are integration problems).
i.e., the possibility of the event {δ_{1} = }.
Optimization as possibility calculus
In the next sections, abusing the notation, an event A will be usually described by a set of constraints on the decision variables δ.
In this way, numeric constrained optimisation problems may be subjectively interpreted in possibilistic terms: the cost J(δ) will be interpreted as the logpossibility of δ and, by definition, unfeasible values of decision variables will be assigned zero possibility.
Let us now review some other relevant definitions and issues in possibilistic calculus.
Necessity
In a binary setting, all solutions belong to a subset A if and only if π(A) = N(A) = 1; there exist solutions in A (and solutions outside A) if π(A) = 1 but N(A) = 0, and there are no solutions in A if π(A) = 0.
Extending the measures π(A), N(A) to [0,1] provides a natural gradation of such concepts: π(A) = 0.95, N(A) = 0.1 would indicate that there are very possible solutions in A, but not all of them are in there (there are solutions with possibility 10.1 = 0.9 outside A).
Interactivity and possibilistic conjunction
The possibilistic analog to statistical independence is the noninteractivity.
which can be read as "the possibility of event A_{1} and event A_{2} is the product of the individual possibilities when the events relate noninteractive variables", interpreting, as usual in literature, set intersection as a linguistic conjunction.
On the following, given individual cost indices J_{1}(δ_{1}), J_{2}(δ_{2}), etc. relating to different constraints, the above expression (8) will be the one used in most cases to define a possibility distribution in the product space. In this way, we are interpreting the possibilistic conjunction operator in [29] as an algebraic product of possibilities, i.e., stating an underlying noninteractivity assumption between different constraints. Note, however, that the interactivity assumption is not always intuitively needed. In the other extreme (total interactivity: variables δ_{1} and δ_{2} fully "correlated", for instance equal), we would have: π(A_{1} ∩ A_{2}) ≤ max(π(A_{1}), π(A_{2})), which would suggest the maximum possibility as the conjunction operator when two events affect exactly the same decision variables. In between those two extremes, other choices may be also possible (T norm operators [30]).
Conditional possibility
The possibilistic analog to conditional probability is conditional possibility.
that is, computing the possibility by subtracting the cost associated to event B from the cost of any of its subsets.
That is, the conditional distribution can be obtained by dividing the possibility distribution function for all points in a set by the maximum possibility of them, i.e., normalising the possibility distribution on a restricted conditioning domain B to a maximum equal to one.
The conditional definitions allow for an analogy to Bayesian inference: if we assume that B is actually certain (whatever the a priori possibility π(B) was), then conditional possibility may be understood as an a posteriori possibility.
Possibility versus probability
Both possibility theory and probability theory are frameworks for handling uncertainty in constraint satisfaction problems. Basically, a subjective interpretation would assign high possibility to events with high probability. Hence, in a first approximation, userdefined probabilities and possibilities should be related by an implicit monotonically increasing function. Possibilitynecessity measures have also been linked to imprecise probabilities [31]. However, once aggregation takes place (via sums and integration in probability, via maximisations in possibility), although the subjective interpretation might be considered similar, there is no longer an implicit function relating probability and possibility. For further discussion on possibility, probability, and other uncertain reasoning frameworks, and their interrelations, the reader is referred to [31–33].
Ideally, probabilistic results would be preferable (to confidently assert that, e.g., 95% of cases a flux estimate will lie in a particular interval). However, there are some drawbacks: (i) exact probabilistic inference under equality and inequality modeling constraints is computationally hard (multivariate integration on irregular sets) (ii) some of the a priori Bayesian probabilities are in practice rough usergiven estimates, (iii) some of the assumptions (linearity of transformation, Gaussian distributions) may not hold in practice, and (iv) there may be some uncertainty in the model parameters or in the model probabilities. Thus, as practical use of probability does not fully adhere to the theoretical assumptions, its results should be interpreted with some flexibility. As this work will discuss, the proposed possibilistic framework is much less demanding computationally (using optimization instead of integrals, so largescale cases become tractable) and gives similar results to the probabilistic approach in realistic cases.
The objective of the next sections is to set up a possibilistic framework for efficient computations in metabolic flux analysis.
Preliminaries for metabolic flux analysis
where v = (v_{1}, v_{2},...,v_{ n })^{ T }is the ndimensional vector of metabolic fluxes.
where e_{ m }represents measurements errors and represents the actually measured flux value. In our approach, the measurement uncertainty is translated into an apriori possibility distribution for e_{ m }from sensor characteristics. Other approaches consider different choices, as discussed below.
where it is assumed that e_{ m }are distributed normally with a mean value of zero and a variancecovariance matrix F.
Since all the constraints are linear equalities, the analytic solution of this minimization problem can be obtained, resulting in the expressions to estimate v_{ u }and v_{ m }that are typically seen in literature (e.g. [11, 36]). However, with this formulation TMFA has some important limitations: (i) irreversibility constraints, or any other inequality constraints, cannot be considered, (ii) measurement errors are assumed to be normally distributed, (iii) it only provides uniquevalued flux estimation, and (iv) it needs a high number of measurable fluxes to be of use – system (14) has to be determined and redundant [37].
Features of Possibilistic MFA and alternative approaches
Feature  TMFA  LSMFA  FSMFA  Monte Carlo  PMFA 

Considers irreversible reaction  x  x  x  x  
Usable in scenarios lacking measurements  x  o  x  
Includes a check of consistency  x      o  x 
Flexible description of measurements errors    x  x  
Richer estimations (not only pointwise)    x  x  
Computational efficiency  x  x  x  x 
In the following sections we introduce a possibilistic framework for MFA that brings several interesting features: (i) it overcomes all the mentioned limitations of TMFA, (ii) has the ability to detect, and handle, inconsistencies between measurements and model, and furthermore (iii) with high computational efficiency.
Results and discussion
Possibilistic MFA: problem statement
In this section the possibilistic framework for MFA flux estimations is discussed. First, we define a set of timeinvariant constraints derived from the metabolism being modelled. Then we incorporate the constraints imposed by the measured fluxes, representing its uncertainty, by means of auxiliar slack decision variables and a cost index. In this way, the notion of "degree of possibility" is incorporated. Finally, it will be shown how (linear) optimisation problems will be able to settle queries about the most possible fluxes, the possibility distributions, etc.
Modelbased constraints
where D is a diagonal nxnmatrix with D_{ i, i }= 1 if the flux i is irreversible (otherwise 0).
Other modelbased constraints can be defined in an analogous way. For instance, elementary balances or degree of reduction balances might be incorporated into (17) as additional constraints [11]. It may be also possible to add constraints based on standard Gibbs free energy changes [43, 44] or extracellular metabolites concentrations [45].
Incorporating the measurements
Estimating the nonmeasured fluxes would amount for solving the above equations (17), where some of the elements in vector v are measured (v_{ m }). However, this simple approach will be impractical in two very common situations:

The measurements are very few, so the system has many possibly infinite solutions.

real measurements do not exactly satisfy the constraints due to measurements (and modelling) errors. Therefore, no solution will be found. (For instance, an unfeasible set results with the constraint v_{1} = v_{2} and the measurements {v_{1} = 0.5, v_{2} = 0.5001}.)
Hence, the approach needs refinements to deal with a lack of measurements and to introduce the "possibility" of sensor errors and imperfect models. As shown below, such difficulties can be overcame by the introduction of slack variables and a cost index, enabling a grading of the different candidate flux vectors as more or less "possible".
Possibilistic description of measurements
with α ≥ 0 and β = 0 (usually, if sensor error is "symmetric", α and β should be defined to be equal).
Recalling the concepts introduced t v_{ m }he preliminaries section, the interpretation of (19) and (20) may be: "v_{ m }= is fully possible; the more v_{ m }differs from , the less possible such situation is". Indeed, the event A = {v_{ m }= } ≡ {ε_{1}  μ_{1} = 0} will be fully possible – as , achieved at ε_{1} = μ_{1} = 0, and then π(A) = e^{0} = 1. On the other hand, the possibility of the event A corresponding to v_{ m }being different from – say, for instance, A = {v_{ m }= + ρ} ≡ {ε_{1}  μ_{1} = ρ} – will be given by . For instance, with a cost index J(δ) = 5ε_{1} + 5μ_{1}, and a measurement = 0.1, the possibility of the actual flux v_{ m }being v_{ m }= 0.2 is e^{5*0.1} = 0.6065 ("quite" possible), and the possibility of v_{ m }= 1.1 is e^{5*1} = 0.0063 ("almost" impossible).
Remark
As explained in a subsequent section, the weights α and β should be defined related to each measurement's "a priori accuracy".
A global cost index
where α and β are row vectors of sensor accuracy coefficients and ε_{ 1 }, ·μ_{ 1 }correspond to stacking in vectors the artificial variables from individual constraints.
The Possibilistic MFA problem
where the decision variables δ are the actual fluxes v = (v_{ u }, v_{ m }), and the slack variables ε_{ 1 }and μ_{ 1 }.
The cost index J reflects the logpossibility of a particular combination of the decision variables, that is, the logpossibility of a particular flux vector v.
Remark
The PMFA will be cast as a linear programming problem; this is the reason why the nonnegative decision variables ε_{1} and μ_{1} were introduced in substitution of e_{ m }. However, it can be formulated using any other optimization framework. For instance, PMFA can be easily cast under a quadratic programming framework. Throughout the paper linear programming will be assumed due to its great computational performance (solvable in polynomial time). This supposes a great advantage when dealing with largescale metabolic networks. Nevertheless, an example using quadratic programming will be described in a subsequent section to point out the flexibility of the PMFA.
Example 1 – Problem statement
Pointwise flux estimations
The simplest outcome of a PMFA problem is a pointwise flux estimation: the minimumcost (maximum possibility) flux vector. This problem can be conveniently cast as the optimisation of a linear functional subject to linear constraints.
being its degree of possibility π(v_{ mp }) = exp(J_{ min }).
The obtained flux vector v_{ mp }contains the most possible flux values compatible (consistent) with the model and the measurements. A possibility equal to one must be interpreted as the flux vector being in complete agreement with the model and the original measurements. Lower values of possibility imply that v_{ mp }corresponds to fluxes v_{ m }deviated from the measurements .
Notice that as π(v_{ mp }) = π( ), it can be interpreted as the "a priori" possibility of encountering the measurements ; so if it is low, this implies that either (a) there is a gross error in the measurements, (b) there is an error in the model, or both. Therefore, the maximum possibility can be used to evaluate consistency and detect errors (inconsistencies between data and models). We will come back to this point in a subsequent section.
Example 1, continued
Consider again the model and the measurements given in Figure 1. The maximum possibility flux vector resulting from (24) is v_{ mp }= (0.75, 9, 30.25, 8.25, 31, 39.3), with a possibility of e^{0.3} = 0.74. The most possible flux vector being not fully possible (peak value not equal to 1) indicates that the measurements and the model are not in complete agreement. Indeed, as a matter of fact, the model says that v_{2}  v_{4} = v_{5}  v_{3}, but = 1 and = 1. Should the measurements had been fully compatible with the constraints imposed by the metabolic network – i.e. = 10, = 30, = 30 and = 10 – the maximum possibility flux vector would have been v_{ mp }= (0, 10, 30,10, 30, 40), with a possibility of π(v_{ mp }c) = 1.
Notice also that the possibility depends on the reliability associated to each measurement. For instance, if all the measurements were supposed to be more reliable – say α' = 10·α and β' = 10·β – the possibility distribution functions would be narrower. The interpretation of the new coefficients would, therefore, be that the same deviation from the fluxes of maximum possibility will be now be considered as a less possible fact.
Possibility distributions as flux estimations
Clearly, a pointwise flux estimation is limited in a situation where multiple flux values might be reasonably possible. To face these situation, marginal and conditional possibility distributions (and intervals) can be obtained, again, by solving linear optimisation problems. These provide a much more informative flux estimation than a pointwise one, such as the maximum possibility flux vector, or the interval of minimummaximum possible values in [38].
Marginal possibility distributions
Marginal possibility distributions (2) can be easily plotted and give a valuable information for the end user: they show, and rank, the possible values for each flux in the network.
Hence, plotting the marginal possibility for a range of fixed given values f (taken within a prespecified range) will provide the marginal possibility distributions that be interpreted as the "distribution of the possible values for each flux in the network, given the measurements" (see Figure 2, left).
This provides a highly efficient procedure to compute a possibility distribution: compute "cuts" of possibilities between 0 and 1, say, 0.1, 0.2, etc. (computing the marginal possibility of all the fluxes in the network by means of a grid of points is linear in the number of grid points and polynomial in the number of fluxes). This approach is better than defining a range of values f and computing its possibility with (25) because it avoids the problem of determining the most convenient step size and bounds of the flux (which, usually, are not known a priori).
Conditional possibility distributions
Remember that conditional distribution can be obtained normalising the marginal possibility distribution to a maximum equal to one (see Figure 2).
Conditional possibility may be understood as an a posteriori possibility: π(v_{ i }= f ) is the possibility of v_{ i }having the value f, if we assume that is actually certain, i.e. that the model and the measurements are correct.
(A posteriori) Possibilistic intervals
The upper bound would be obtained by replacing minimum by maximum.
These possibilistic intervals have a similar interpretation to "confidence intervals" ("credible intervals") in Bayesian statistics, providing a concise flux estimation that can be represented by means of a boxplot chart (see Figure 2, right).
Example 1, continued
Notice too that the uncertainty on the measurements is often strikingly reduced through the flux estimation. For instance, the estimation of v_{4} – whose measurement was quite unreliable a priori – has been significantly improved, once model constraints and other measurements have been incorporated. This reflects the already noticed fact that the metabolic network structure greatly constrains the possible values of fluxes for a given, typically small, set of measured flux values. The plots of marginal possibility can also detect multiple flux vectors with maximum possibility (possibility distribution functions with flat top). Figure 3B depicts the maximum possibility flux estimation and three possibilistic intervals by means of a boxplot chart. The intervals point out that, for instance, the highly possible a posteriori values of v_{5} are those in [30.75, 31] (possibility greater than 0.9) and that those in [29.5, 32] are also quite possible (possibility greater than 0.5), while those outside [27, 34.5] are almost impossible as their a posteriori possibility is lower than 0.1.
Possibilistic MFA: Refinements
Now that the basics of the PMFA framework have been introduced, some refinements will be discussed.
A better description of measurement's uncertainty
The formulation used above to describe the uncertainty of the experimental measurements might be considered somehow limited in some applications.
Fortunately, it is very easy to add new slack variables, and modify the (23) and the cost index (22), allowing to work with possibility distribution functions of different characteristics.
The possibility of is one and the possibility of the actual flux being v_{ m }being out of the referred interval depends on the cost index weights (α and β).
For instance, a band with possibility equal to one can be used to account for systemic errors in measuring a particular flux, and a couple of additional slack variables may be defined to account for the decreasing possibility of random errors. These kind of representation of measurement uncertainty will be illustrated in subsequent examples.
Remark
Notice that more slack variables can be added to achieve a more complex representation of the measured flux uncertainty. In fact, any convex representation of the logpossibility uncertainty can be approximated if a sufficient number of slack variables are incorporated (ε_{1}, μ _{2}, ε_{2}, μ_{2},...). Details are omitted for brevity.
Considering uncertainty in the model structure
Until now, the modelbased constraints (23) – the stoichiometric relationships, reaction's irreversibility, or any other – have been considered as hard constraints; only those flux vectors v that exactly satisfy them could be feasible solutions. However, these constraints can be "softened" via suitable slack variables to consider uncertain knowledge. Then, these additional slack variables may be used in a cost index to generate a possibility distribution.
with ϵ and ν being slack variables penalised in an optimisation index J = f(ϵ, ν), typically with linear cost index terms, γ ϵ + τυ, in an analogous way to the discussion on uncertain measurements.
Notice also that a "softened" inequality restriction is nothing but an equality one with no penalisation on one of the slack variables above. For instance a ≤ b + ε can be expressed as a = b + ε  μ with free μ.
Such softened model constraints may be used to roughly incorporate imprecision in the model arising, for instance, from noncompliance with the pseudosteadystate assumption, partial unbalance of some metabolites or uncertain yields. Although these issues require further research, let us outline some preliminary ideas below.
Relaxing the pseudosteady state assumption
Equation (13) derives from the dynamic mass balance around the internal metabolites c, where it is assumed that ≈ 0 and that the term μ·c is negligible (μ denotes the growth rate). Adding slack decision variables to (13), as it has been explained, makes it possible to relax this assumption.
Partial unbalance of metabolites
Sometimes, a metabolite cannot be assumed to be balanced because there are reactions producing or consuming this metabolite that have not been taken into account in the network; for instance, this is often the case for the cofactors, ATP, NADP, etc. This unknown consumption/production can be represented by means of slack variables (e.g. ϵ and υ) if some interval limits (e.g. ϵ_{ max }and υ_{ max }) are provided.
Uncertainty in stoichiometric yields
However, if the flux v_{ r }is reversible, inequalities in (33) cannot be set up, and the approach is no longer applicable. Integrating modal interval arithmetic [46] in the proposed framework might be a possible option, under research at this moment.
Illustrative examples of other features of PMFA
Other features of Possibilistic MFA (PMFA) will be briefly illustrated using the simple metabolic network in Figure 1.
Example 2: Errors detection in measurements and model
As earlier mentioned, the value of the peak possibility in the resulting flux distribution provides an indication of the agreement between the model ( ) and the measurements ( ). A low degree of possibility means that the model and the measurements are inconsistent. That is, that there is not any flux vector "near" the measured values satisfying the modelbased constraints. Therefore, if the the maximum possibility flux vector has a low value, one must assume that either (a) there is an important error in one or more measurements, (b) there is a relevant error in the model (e.g. a mass balance is not closed, or a metabolite is not at steady state), or both.
If a high inconsistency (low possibility) is detected, it is possible to investigate what is causing it, and thus correct the measurements or improve the model. Following a straight approach, we can remove one measured flux at a time and perform the flux estimation to determine if the removed measurement was causing the low possibility. If this is the case, we may consider the following alternatives: (a) consider that is a totally unreliable measurement and, thus, accept the flux estimation inferred from the others measurements, (b) obtaining either again, or a different measurable flux which could provide additional information, (c) consider a reliable piece of data and, hence, conclude that there is an error in the model or its assumptions. In case (c), a similar approach can be used to investigate which particular modelbased constraint is causing the low possibility – by "softening" the suspicious constraints one at a time.
Example 3: Scenario lack of data
One of the features of PMFA is that it can be used even if there is a lack of measurements; i.e. even if (14) is underdetermined or not redundant [37].
Example 4: Using quadratic programming
To show how PMFA can be cast within other optimisation frameworks, an example using quadratic programming will be discussed. We define as = v_{ m }+ e_{ m }and J = ·W·e_{ m }, where W is a diagonal matrix of weights. Hence, we have for each measurement, i.e. measurements uncertainty is represented as a quadratic possibility distribution.
Let us continue with our example using the measurements of Figure 1, but representing them with the quadratic formulation just introduced. The original possibility distribution of single measurements (dashed lines) and the possibility distributions computed with PMFA (solid lines) are depicted in the Figure 5. Notice that results are similar to those obtained in the previous example (Figure 1), where the standard linear programming framework was used (even if additional auxiliar variables ϵ_{2}, μ_{2}, etc. were not used). However, the qualitative similarity between the results makes the author think that, in most cases, the linear programming setup is expressive enough and much more efficient than quadratic or other more complex optimization cases.
Example 5: Comparison with other methods
This example compares PMFA with traditional MFA and some of its extensions. We continue using the network depicted in Figure 1, and perform the estimations with PMFA, Traditional MFA (TMFA), MFA as a constraint, leastsquares problem (LSMFA) and the flux spectrum (FSMFA).
Inconveniently, errors have to be approximated with a normal distribution so that TMFA and LSMFA can be used (see preliminaries). For the estimations with FSMFA we represent the measurements with the interval of 95%, or 2σ (see [38]). All the results are depicted in Figure 6. Notice that TMFA assigns a negative value to an irreversible flux, v_{1}, since it is not taking these constraints into account – this was predictable, but it must be highlighted because TMFA is still being widely used in the literature. The results also point out that the possibilistic distribution (and intervals) are much more informative than the pointwise estimations of TMFA and LSMFA, or the intervals of FSMFA. Basically, pointwise estimations fail when several flux values reasonably possible, whereas the flux spectrum interval tend to be conservative. Furthermore, remember that TMFA and LSMFA cannot be used in scenarios lacking data, such as example 3, where PMFA was shown to be valuable.
To complete this perspective, the next section will discuss a comparison between PMFA and Monte Carlo approaches.
Example 6: Comparison with Monte Carlo
Notice also that this is a simple case where Monte Carlo can be applied. Nonetheless, its worst performance is clear: the cost of computing the possibility distributions is polynomial in the number of fluxes (as shown above), whereas the cost of a Monte Carlo approach grows exponentially with the number of independent decision variables.
Largerscale Example: C. glutamicum
In this section we apply Possibilistic MFA (PMFA) to a mediumsize example. For illustrative purposes, we have chosen a very wellknow metabolic model of Corynebacterium glutamicum.
Metabolic network model
The metabolic network of C. glutamicum has been taken from [47] and is a slight variation of the one originally constructed in [48, 49]. The reactions considered in describing the biochemistry of the primary metabolism of C. glutamicum necessary to support lysine and biomass synthesis from glucose are given in the additional file 1. A reaction of ATP dissipation is included in the network, so that the ATP balance could be maintained, without actually constraining the flux space. On the contrary, the cofactors NADP, NAD and FAD are supposed to be balanced. The reaction for biomass formation is an approximation using as reactants those amino acids that explicitly appear in the network and the precursors of the other amino acids synthesized by C. glutamicum.
PMFA setting
The stoichiometric relationships, embedded in a 36 × 40 stoichiometric matrix, and the irreversibility of certain reactions, embedded in a 40 × 40 diagonal matrix, define our modelbased constraints ( ) according to (17). Both matrices are given in the additional file 1.
Experimental measurements
PMFA setting
Using the data in Figure 8, we have built a possibilistic representation of single measurements defining convenient auxiliar variables and weights. The criterion to choose the weights was: full possibility for v_{ m }∈ ± σ/2 and possibility 0.5 for those in ± σ. The values in ± 2·σ have possibility 0.1 (σ denotes standard deviation. If errors are assumed to be normally distributed, these levels correspond to the probabilistic confidence intervals of 38%, 68% and 95%, respectively). The resultant decision variables and weights define our measurementbased constraints ( ) according to (21). These possibilistic representations are depicted in Figure 8.
Possibilistic flux estimation of C. glutamicum
We used all the available measurements – v_{ GLC }(1), v_{O 2}(34), v_{NH 3}(35), v_{ LY }(37), v_{ Thre }(38), v_{CO 2}(39) and v_{ Bio }(36) – to obtain the maximum possibility flux vector (results given in the additional file 1). The flux vector had a degree of possibility 0.38, which could be considered "low" if one considers that a significant degree of uncertainty was already being taken into account (table 1). We then obtained the marginal possibility distributions for each flux, which inspection indicated that the low possibility was almost completely caused by only one measured flux, v_{NH 3}(35). This suggests that this measurement was inaccurate, or that its standard deviation was underestimated. Interestingly, this flux was indeed the most uncertain one in the original dataset (its standard deviation was a huge 44 mM/h for a nominal value of 64.8 nM/h).
Possibilistic flux estimation lacking measurements
Possibilistic flux estimation with uncertain model
As explained above, we can "soften" the modelbased constraints to relax the pseudosteady state assumption. As example, we assumed a degree of uncertainty around all the mass balances introducing decision variables ϵ_{ 1 }and υ_{ 1 }and weights γ_{1} = τ_{1} = 2 (see Figure 8). Hence, flux vectors which imply small accumulations of some metabolites will be accepted, yet considered less possible.
It could be also stated that the metabolic network used herein, the one introduced by Vallino et al., relies on an unrealistic assumption: that cofactors NADP, NAD and FAD are balanced [50, 51]. To avoid this, we can remove these metabolites from our stoichiometric matrix or, as an alternative, use the expressivity of the possibilistic framework to allow a certain degree of unbalance for these metabolites. Just as example, herein we assumed that cofactors may be unbalanced with some limits (say, 30 mM/h for NADP/NADPH and 15 mM/h for FAD/FADH and NAD/NADH). This "knowledge" can be easily incorporated into our model defining the convenient auxiliar variables and weights (as explained above).
Conclusion
In this paper we have discussed a possibilistic framework for the estimation of the metabolic fluxes shown by cells at given conditions given.
Considering ordinary constraintsatisfaction problems, metabolic fluxes fulfilling a set of modelbased constraints and compatible some experimental measurements are "possible", otherwise "impossible". In this paper, this idea is refined to cope with uncertain knowledge – in the form of measurements errors or imperfect models – by introducing the notion of "degree of possibility", which enables grading the candidate flux values as more or less possible. Then, possibilistic MFA is able to query the flux vector of maximum possibility. Moreover, when multiple flux vectors might be reasonably possible, the marginal and conditional possibility distributions for each flux can be computed.
Possibilistic MFA overcomes several limitations of traditional MFA and some of its extensions. It considers measurements uncertainty in a flexible way (e.g. nonsymmetric error or a band of uncertainty due to systemic error) and also model imprecision, and it is reliable even if only a few fluxes are measurable (a common scenario). Possibilistic MFA also computes possibility distributions (and intervals) which are more informative than pointwise estimations when multiple flux values might be reasonably possible. These distributions are also better than the intervals provided by the flux spectrum, or other methods giving upper and lower bounds for the fluxes. In addition, Possibilistic MFA has the ability to detect, and handle, inconsistencies between measurements and model. Finally, it must be remarked that Possibilistic MFA estimations have been cast as linear optimisation problems, for which widelyknown and efficient tools exist (a MATLAB script solving example 1 is given in the additional file 2 to illustrate this point). This great computational performance makes the methodology capable of dealing with largescale or even genomescale metabolic networks.
It must be noticed that there is a challenge when estimating the fluxes in largescale networks because there may be diffierent flux vectors compatible with the few available measurements [52]. Interestingly, the proposed methodology is still of use in this situation: possibilistic MFA will detect all these flux vectors that are equally possible (or even similarly possible) and depict them by means of possibilistic distributions or intervals (e.g. example 3). Unfortunately, if there is a wide range of candidates, the estimation may be sometimes little informative (but at least we can be sure that it is reliable, because all the flux vectors compatible with model and measurements are captured). One strategy to face this difficulty consists of using a rational hypothesis to promote certain flux vector among those that are equally possible. For instance, it can be assumed that cell behaviour has evolved to be optimal in some sense, so that the fluxes are optimally regulated depending on the given environmental conditions, and then invoke this principle to choose particular flux vectors [3, 7, 9]. There might be still alternate optima, but the approach will reduce the range of possible flux vectors. Notice that this optimality principle, or any other hypothesis, might be incorporated into the possibilistic framework as far as they are encoded in in the form of a cost index (but sometimes it will not be the case or, in other cases, its optimization will not be computationally simple). This point requires further work.
Further extension may also address the adaptation of the ideas introduced herein to metabolic flux analysis with data from labelling experiments (13CMFA) [12–14]. Extracellular dynamics could be also taken into account incorporating measurements in different time instants [38]. Finally, we are currently developing a software that implements the Possibilistic MFA methods and its future extensions, which will be freely available for academia.
In summary, this papers introduces a unifying framework for flux estimation and (possibilistic) evaluation of consistency that is flexible, usable in scenarios lacking data, highly informative, and computationally efficient. In our opinion, the combination of computational efficiency and flexibility of the assumptions is a distinctive advantage with respect to other approaches which either may rely on stronger assumptions (chisquared distributions, intervalonly descriptions, absence of irreversibility), or be only databased (so they do not incorporate, say, stoichiometric model balances), or provide only pointwise estimates (of flux or consistency), or be computationally intensive (multivariate integration in a general Bayesian estimation problem).
Declarations
Acknowledgements
This research has been partially supported by the Spanish Government (1st and 3rd authors are grateful to grant DPI200806880C0301, 2nd author is grateful to grants DPI200806731C0201 and PROMETEO/2008/088). FLL is recipient of a fellowship from the Spanish Ministry of Science and Innovation (FPU AP20051442).
Authors’ Affiliations
References
 Palsson B: The challenges of in silico biology. Nature biotechnology. 2000, 18 (11): 114750. 10.1038/81125View ArticlePubMedGoogle Scholar
 Kitano H: Computational systems biology. Nature. 2002, 420 (6912): 20610. 10.1038/nature01254View ArticlePubMedGoogle Scholar
 Palsson B: Systems Biology: Properties of Reconstructed Networks. 2006, Cambridge University Press New York, NY, USAView ArticleGoogle Scholar
 Llaneras F, Picó J: Stoichiometric modelling of cell metabolism. J Biosci Bioeng. 2008, 105: 111. 10.1263/jbb.105.1View ArticlePubMedGoogle Scholar
 Price ND, Papin JA, Schilling CH, Palsson BO: Genomescale microbial in silico models: the constraintsbased approach. Trends in Biotechnology. 2003, 21 (4): 1629. 10.1016/S01677799(03)000301View ArticlePubMedGoogle Scholar
 Kauffman KJ, Prakash P, Edwards JS: Advances in flux balance analysis. Current Opinion in Biotechnology. 2003, 14 (5): 4916. 10.1016/j.copbio.2003.08.001View ArticlePubMedGoogle Scholar
 Schuetz R, Kuepfer L, Sauer U: Systematic evaluation of objective functions for predicting intracellular fluxes in Escherichia coli. Mol Syst Biol. 2007, 3: 119 10.1038/msb4100162PubMed CentralView ArticlePubMedGoogle Scholar
 Edwards JS, Ibarra RU, Palsson BO: In silico predictions of Escherichia coli metabolic capabilities are consistent with experimental data. Nature biotechnology. 2001, 19 (2): 12530. 10.1038/84379View ArticlePubMedGoogle Scholar
 Schilling CH, Covert MW, Famili I, Church GM, Edwards JS, Palsson BO: Genomescale metabolic model of Helicobacter pylori 26695. J Bacteriol. 2002, 184 (16): 458293. 10.1128/JB.184.16.45824593.2002PubMed CentralView ArticlePubMedGoogle Scholar
 Schuster S, Pfeiffer T, Fell DA: Is maximization of molar yield in metabolic networks favoured by evolution?. J Theor Biol. 2008, 252 (3): 497504. 10.1016/j.jtbi.2007.12.008View ArticlePubMedGoogle Scholar
 G Stephanopoulos AA, Aristidou A: Metabolic Engineering: Principles and Methodologies. 1998, Academic Press, San Diego, USAGoogle Scholar
 Sauer U: Metabolic networks in motion: 13Cbased flux analysis. Mol Syst Biol. 2006, 2: 62 10.1038/msb4100109PubMed CentralView ArticlePubMedGoogle Scholar
 Szyperski T: 13CNMR, MS and metabolic flux balancing in biotechnology research. Q Rev Biophys. 1998, 31: 41106. 10.1017/S0033583598003412View ArticlePubMedGoogle Scholar
 Wiechert W: 13C metabolic flux analysis. Metabolic Engineering. 2001, 3 (3): 195206. 10.1006/mben.2001.0187View ArticlePubMedGoogle Scholar
 Nookaew I, Meechai A, Thammarongtham C, Laoteng K, Ruanglek V, Cheevadhanarak S, Nielsen J, Bhumiratana S: Identification of flux regulation coefficients from elementary flux modes: A systems biology tool for analysis of metabolic networks. Biotechnol Bioeng. 2007, 97 (6): 153549. 10.1002/bit.21339View ArticlePubMedGoogle Scholar
 Poolman MG, Venkatesh KV, Pidcock MK, Fell DA: A method for the determination of flux in elementary modes, and its application to Lactobacillus rhamnosus. Biotechnol Bioeng. 2004, 88 (5): 60112. 10.1002/bit.20273View ArticlePubMedGoogle Scholar
 Schwartz JM, Kanehisa M: Quantitative elementary mode analysis of metabolic pathways: the example of yeast glycolysis. BMC Bioinformatics. 2006, 7: 186 10.1186/147121057186PubMed CentralView ArticlePubMedGoogle Scholar
 Dubois D, Prade H: Fuzzy relation equations and causal reasoning. Fuzzy Sets and Systems. 1995, 45 (2): 119134. 10.1016/01650114(95)00105T.View ArticleGoogle Scholar
 Sala A, Albertos P: Fuzzy systems evaluation: The inference error approach. Systems, Man and Cybernetics, Part B, IEEE Transactions on. 1998, 28 (2): 268275. 10.1109/3477.662768.View ArticleGoogle Scholar
 Sala A, Albertos P: Inference error minimisation: fuzzy modelling of ambiguous functions. Fuzzy Sets and Systems. 2001, 121: 95111. 10.1016/S01650114(99)001748.View ArticleGoogle Scholar
 Sala A: Encoding Fuzzy Possibilistic Diagnostics As A Constrained Optimisation Problem. Information Sciences. 2008, 178: 42464263. 10.1016/j.ins.2008.07.017.View ArticleGoogle Scholar
 Kumar V, et al.: Algorithms for constraintsatisfaction problems: A survey. AI magazine. 1992, 13: 3244.Google Scholar
 Russell S, Norvig P: Artificial Intelligence: a modern approach. 2003, PrenticeHall, 2Google Scholar
 Jensen F: Introduction to Bayesian networks. 1996, SpringerVerlag New York, Inc. Secaucus, NJ, USAGoogle Scholar
 Hand D: Statistical reasoning with imprecise probabilities. Applied Statistics. 1993, 42: 237238. 10.2307/2347427.View ArticleGoogle Scholar
 Yager R: An introduction to applications of possibility theory. Human Systems Management. 1983, 3: 246269.Google Scholar
 Dubois D, Prade H: Possibility theory – an approach to computerized processing of uncertainty. New York, USA. 1988Google Scholar
 Zadeh L: Possibility theory and soft data analysis. Mathematical Frontiers of Social and Policy Sciences. Edited by: Cobb L, Thrall R. 1981, 69129. Boulder: Westview PressGoogle Scholar
 Dubois D, Fargier H, Prade H: Possibility theory in constraint satisfaction problems: handling priority, preference and uncertainty. Applied Inteligence. 1996, 6 (4): 287309. 10.1007/BF00132735.View ArticleGoogle Scholar
 Benferhat S, Dubois D, Prade H: Syntactic Combination of Uncertain Information: A Possibilistic Approach. Lecture notes in computer science. 1997, 3042. full_text.Google Scholar
 Dubois D, Prade H: Intervalvalued fuzzy sets, possibility theory and imprecise probability. Proceedings of International Conference in Fuzzy Logic and Technology. 2005Google Scholar
 Klir G, Parviz B: ProbabilityPossibility Transformations: a Comparison. International Journal of General Systems. 1992, 21 (3): 291310. 10.1080/03081079208945083.View ArticleGoogle Scholar
 Dubois D, Prade H: Possibility Theory, Probability Theory and MultipleValued Logics: A Clarification. Annals of Mathematics and Artificial Intelligence. 2001, 32: 3566. 10.1023/A:1016740830286.View ArticleGoogle Scholar
 Heijden T, Luyben K: Linear Constraint Relations in Biochemical Reaction Systems: I. Classification of the Calculability... Biotechnol Bioeng. 1994, 43: 110. 10.1002/bit.260430102View ArticleGoogle Scholar
 Heijden T, Luyben M: Linear Constraint Relations in Biochemical Reaction Systems: II. Diagnosis and Estimation of Gross errors. Biotechnol Bioeng. 1994, 43: 1120. 10.1002/bit.260430104View ArticlePubMedGoogle Scholar
 Gambhir A, Korke R, Lee J, Fu PC, Europa A, Hu WS: Analysis of cellular metabolism of hybridoma cells at distinct physiological states. J Biosci Bioeng. 2003, 95 (4): 31727.View ArticlePubMedGoogle Scholar
 Klamt S, Schuster S, Gilles ED: Calculability analysis in underdetermined metabolic networks illustrated by a model of the central metabolism in purple nonsulfur bacteria. Biotechnol Bioeng. 2002, 77 (7): 73451. 10.1002/bit.10153View ArticlePubMedGoogle Scholar
 Llaneras F, Picó J: A procedure for the estimation over time of metabolic fluxes in scenarios where measurements are uncertain and/or insufficient. BMC Bioinformatics. 2007, 8: 421 10.1186/147121058421PubMed CentralView ArticlePubMedGoogle Scholar
 Llaneras F, Picó J: An interval approach for dealing with flux distributions and elementary modes activity patterns. J Theor Biol. 2007, 246 (2): 290308. 10.1016/j.jtbi.2006.12.029View ArticlePubMedGoogle Scholar
 Wiechert W, Möllney M, Petersen S, de Graaf AA: A universal framework for 13C metabolic flux analysis. Metabolic Engineering. 2001, 3 (3): 26583. 10.1006/mben.2001.0188View ArticlePubMedGoogle Scholar
 Kadirkamanathan V, Yang J, Billings SA, Wright PC: Markov Chain Monte Carlo Algorithm based metabolic flux distribution analysis on Corynebacterium glutamicum. Bioinformatics. 2006, 22 (21): 26812687. 10.1093/bioinformatics/btl445View ArticlePubMedGoogle Scholar
 Schmidt K, Nørregaard LC, Pedersen B, Meissner A, Duus JO, Nielsen JO, Villadsen J: Quantification of intracellular metabolic fluxes from fractional enrichment and 13C13C coupling constraints on the isotopomer distribution in labeled biomass components. Metabolic Engineering. 1999, 1 (2): 16679. 10.1006/mben.1999.0114View ArticlePubMedGoogle Scholar
 Henry CS, Broadbelt LJ, Hatzimanikatis V: Thermodynamicsbased metabolic flux analysis. Biophys J. 2007, 92 (5): 1792805. 10.1529/biophysj.106.093138PubMed CentralView ArticlePubMedGoogle Scholar
 Feist AM, Henry CS, Reed JL, Krummenacker M, Joyce AR, Karp PD, Broadbelt LJ, Hatzimanikatis V, Palsson BØ: A genomescale metabolic reconstruction for Escherichia coli K12 MG1655 that accounts for 1260 ORFs and thermodynamic information. Mol Syst Biol. 2007, 3: 121 10.1038/msb4100155PubMed CentralView ArticlePubMedGoogle Scholar
 Mo ML, Palsson BØ, Herrgård MJ: Connecting extracellular metabolomic measurements to intracellular flux states in yeast. BMC Syst Biol. 2009, 3: 37 10.1186/17520509337PubMed CentralView ArticlePubMedGoogle Scholar
 Gardeñes E, Sainz M, Jorba L, Calm R, Estela R, Mielgo H, Trepat A: Modal Intervals. Reliable Computing. 2001, 7 (2): 77111. 10.1023/A:1011465930178.View ArticleGoogle Scholar
 Gayen K, Venkatesh KV: Analysis of optimal phenotypic space using elementary modes as applied to Corynebacterium glutamicum. BMC Bioinformatics. 2006, 7: 445 10.1186/147121057445PubMed CentralView ArticlePubMedGoogle Scholar
 Vallino J: Identification of branchpoint restrictions in microbial metabolism through metabolic flux analysis and local network perturbations. PhD thesis. 1991, Massachusetts Institute of Technology, Dept. of Chemical EngineeringGoogle Scholar
 Vallino JJ, Stephanopoulos G: Metabolic flux distributions in Corynebacterium glutamicum during growth and lysine overproduction. Reprinted from Biotechnology and Bioengineering, Vol. 41, Pp 633–646 (1993). Biotechnol Bioeng. 2000, 67 (6): 87285. 10.1002/(SICI)10970290(20000320)67:6<872::AIDBIT21>3.0.CO;2XView ArticlePubMedGoogle Scholar
 Yang TH, Wittmann C, Heinzle E: Respirometric 13C flux analysisPart II: in vivo flux estimation of lysineproducing Corynebacterium glutamicum. Metabolic Engineering. 2006, 8 (5): 43246. 10.1016/j.ymben.2006.03.001View ArticleGoogle Scholar
 Marx A, de Graaf AA, Wiechert W, Eggeling L, Sahm H: Determination of the fluxes in the central metabolism of Corynebacterium glutamicum by nuclear magnetic resonance spectroscopy combined with metabolite balancing. Biotechnol Bioeng. 1996, 49 (2): 11129. 10.1002/(SICI)10970290(19960120)49:2<111::AIDBIT1>3.0.CO;2TView ArticlePubMedGoogle Scholar
 Bonarius H, Schmid G, Tramper J: Flux analysis of underdetermined metabolic networks: the quest for the missing constraints. Trends in Biotechnology. 1997Google Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.