Adaptive bi-level programming for optimal gene knockouts for targeted overproduction under phenotypic constraints

Backgrounds: FBA and MOMA

Before introducing our new bi-level programming problem to identify optimal metabolic genes or reactions to delete for the maximization of targeted bio-productions, we first review the mathematical foundations of FBA [14] and MOMA [11]. FBA provides appropriate simplifications for metabolic flux analysis by assuming the balance of production and consumption fluxes at steady states of metabolic network models. Specifically, with the prior stoichiometry knowledge, FBA assumes that the weighted sum of network fluxes based on stoichiometric coefficients S is 0: ${\sum }_{j=1}^M{S}_{ij}{v}_j=0$, 1 ≤ i ≤ N, in which we assume that the network model has M reactions and N metabolites in total; S _ij is the stoichiometric coefficient of metabolite i in reaction j; and v _j denotes the flux value of reaction j. For wild-type strains, as mentioned above, a common assumption is that their steady-state flux values follow an optimal distribution that maximizes the biomass production rate. The steady-state flux distribution is approximately solved as a LP problem to maximize the biomass production flux: ${\mathsf{\text{max}}}_{v_{j,1\le j\le M}}$v _biom subject to the FBA stoichiometry constraints, in which v _biom is defined by summing up the metabolite precursors that contribute to the biomass production in FBA [11]. In OptKnock, the optimal gene knockout strategy is to remove genes or reactions by setting the corresponding v _j to zero with the resulting knockout flux distribution maintaining biomass maximization assumption.

where v _j represents the flux value of reaction j in mutant strains and w _j is the corresponding flux value in wild-type strains. The flux value for biomass production v _biom is similarly defined as mentioned earlier. In addition, the glucose flux value v _glc denotes the glucose consumption rate, which is often set to a fixed value v _{glc_uptake} . Finally, $v_j^{min}$ and $v_j^{max}$ are the lower bound and upper bound for v _j , which are determined by the availability of nutrients or the maximal fluxes that can be supported by enzymatic pathways [11].

New bi-level programming framework

Following the modeling strategy in OptKnock [13], we aim to derive optimal gene knockout strategies, which consequently remove corresponding reactions for desired biomedical overproduction while maintaining obligatory cellular conditions, for example, cell mortality. However, as it has been shown that the assumption of biomass maximization for steady-state cellular conditions may not correctly predict metabolic flux distributions for knockouts [11, 13], we replace the internal cellular objective of maximizing biomass yield in OptKnock [13] by the MOMA assumption [11], which has led to better predictions of steady-state flux allocations for genetically engineered strains. With this critical change from OptKnock, we formulate a novel bi-level programming model for gene knockouts in which the inner optimization problem is a QP problem.

in which K is the allowed maximum number of knockouts and v _chemical corresponds to the reaction that produces the desired biochemical production target. Note that we do not count in the flux change for the target reaction in the inner problem as it would contradicts to our primal optimization for maximal biochemical overproduction.

Adaptive linearization strategy for an exact optimal solution

We emphasize that the nested inner optimization problem is a QP problem with respect to flux allocation v _j in knockout strains. As this nested inner problem is convex, we can still get its dual problem and the strong duality condition still holds for the inner primal and dual problems. Following the similar direction of [13], we can develop a single-level equivalent formulation by enforcing the objective value of the inner primal problem equal to that of its dual problem. However, the resulting formulation will be a mixed integer quadratically constrained programming problem, which poses a huge computational challenge when solving real problems. Because of this major change due to the introduction of the inner QP problem under the MOMA assumption, the transformation in OptKnock to a typical single-level mixed integer linear programming (MILP) problem based on the linear programming (LP) duality theory is not directly applicable any more.

To derive efficient solution algorithms for our new bi-level programming gene knockout problem, we adopt a novel adaptive linearization solution strategy to tackle the computational complexity introduced by the inner QP problem. Specifically, we propose to adaptively represent the quadratic terms in the objective function of the inner problem using a set of linear functions as illustrated in Figure 1(A), which yields a LP approximation for the nested inner problem. With a given piecewise linearization of the inner problem, we can convert our new bi-level model into a single-level problem based on the LP strong duality. For the linearized problem, we can obtain the optimal solution similarly as in [13] by solving the transformed single-level MILP problem. In order to obtain the exact optimal solution to the original bi-level problem with the inner QP problem, we adaptively create necessary pieces on the fly to approximate the quadratic objective function until the solution converges.

The basic idea of adaptive piecewise linearization is illustrated in Figure 1(B-D). We denote the initial starting solution by v₁, which can be represented by a convex combination of endpoints of piecewise segments for a given piecewise linearization. The corresponding quadratic objective function value at v₁ is denoted by M₁, which can be approximated linearly by $M_1^{\prime }$as the convex combination of the corresponding objective function values at segment endpoints A, B, C and D. We iterate the procedures to solve the linearized single-level MILP problem and to adaptively add piecewise linear segments to better approximate the inner quadratic objective function as illustrated in Figure 1(B-D) until the optimal solution of the MILP problem achieves the desired precision with respect to the approximation of the inner QP objective function. This adaptive linearization strategy has the guarantee that the final solution converges to the exact optimal solution. More importantly, it is much more efficient than directly solving mixed integer quadratic constrained problem without linearization and hence it allows us to solve for large-scale metabolic networks.

With this basic understanding of our new bi-level model and adaptive piecewise linearization solution strategy, we describe the detailed algorithm in the following sections.

Piecewise linearized inner problem

Here, both w _j and $v_j^t$ are constants and we have removed the constant terms $w_j^2$ in the original objective function. This linear approximation of the original inner objective function based on the MOMA criterion now enables the solution strategy to the bi-level programming problem by taking advantage of the LP strong duality property [16], for which the objective function values for the primal and dual problems of the approximated inner LP problem must be equal to each other at optimality if both of them are bounded. With this duality condition, the bi-level programming problem can be solved as a single-level MILP problem by including the dual problem formulation and enforcing that the primal and dual problems share the same objective function value as in [13].

This final single-level MILP problem can be solved effectively by professional solvers, such as CPLEX [17]. We note that our new MOMA-based knockout optimization problem has a larger problem size with a larger number of variables and constraints as multiple linear functions are used to approximate the inner quadratic function.

Adaptive strategy

We have shown that we can effectively solve the linearized bi-level programming problem in the previous section. However, due to the linearization of the original quadratic MOMA objective function, the obtained result for a given linearization scheme is an approximate solution but not exact. In addition, the closeness to the exact optimal solution is directly determined by the number of segments for each flux to approximate the quadratic function $v_j^2$. In order to obtain the exact optimal solution to the original bi-level programming problem, we adopt an adaptive strategy, in which piecewise linearization is implemented adaptively from the coarse to fine levels. As the original inner problem is to minimize the quadratic MOMA objective function, which is convex. It is easy to prove that the approximate optimal solution for a given linearization will have each flux v _j fall within one segment. In other words, for each flux v _j , piecewise variables ${\beta }_j^t$ only have either one (at endpoints) or two adjacent non-zero values for the approximate solution as illustrated in Figure 1.

Based on the differences and the state of vector β _j for all flux values, we adaptively add new piecewise linear segments to better approximate the corresponding contributions from each reaction flux to the quadratic objective function in the inner problem. By repeating the above procedure as shown in Figure 1(B-D), we can iteratively solve the problem by adaptively improve the piecewise linearization from coarse to fine levels until adding pieces does not change the objective value. If every Δ _j is less than a very small number ϵ and every maximum value in β _j is larger than a constant number θ that is close to 1, we can say the algorithm has converged. To speedup the algorithm, the knockouts from previous iteration are used to get a low bound for the MILP problem. Algorithm 1 provides the pseudo code for our adaptive linearization solution strategy to identify optimal knockout strategy for biochemical overproduction under the MOMA constraint.

Algorithm 1 Adaptive bi-level MOMAKnock.

Initialize variables.

Initialize the piecewise linearization with k pieces

repeat

   Solve the inner primal problem based on previous knockouts to get a low bound objL;

   Solve the MILP problem with the low bound objL;

   for Each flux j do

      Compute Δ _j .

      if Δ _j >ϵ or $\underset{t}{\mathsf{\text{max}}}{\beta }_j^t<\theta$then

         Add a segment point at $v_j^{t^{*}}{\beta }_j^{t^{*}}+v_j^{t^{*}+1}{\beta }_j^{t^{*}+1}$; (${\beta }_j^{t^{*}}$ and ${\beta }_j^{t^{*}+1}$ are nonzero)

      end if

      end for

until Added segments do not improve the objective function

Succinate production on AntCore metabolism network

Based on the results from OptKnock in Table 1 we can see that the objective function values for the targeted succinate production are indeed high with the biomass maximization assumption as constraints. For example, when the knockout number K = 2, OptKnock can achieve as high as over 72.44 percent of the theoretical maximum succinate flux value 142.16 mmol/gDW · hr for its optimal solution. However, when we evaluate the actual flux values under the MOMA objective, the resulting succinate flux value drops to as low as 18.51 percent. Similarly, for K = 3 and 5, OptKnock also derives high succinate flux values under the biomass maximization assumption while the actual values drop significantly in suggested knockout strains under the MOMA objective. When K = 4, removing four reactions leads to the optimal succinate flux value at 118.71 mmol/gDW · hr. The suggested knockout strategies maintain to obtain a high value as high as 84.56 mmol/gDW · hr for succinate production in the MOMA flux distribution. However, we notice that the corresponding biomass flux values in both OptKnock and MOMA flux distributions are at 5.00 mmol/gDW · hr, which is the minimum biomass flux value set in our experiments to guarantee living cells. Hence, we believe that the derived knockout strain may not be robust, which does not lead to practically feasible knockout strategies but causes the death of cells. We investigate the suggested knockout reactions when K = 3 and 4 as the MOMA biomass flux value when K = 3 reaches 5.23 mmol/gDW · hr, close to the minimum value. When K = 3, the most important Transhydrogenation reaction (nadh → nadph) that produce nadph (Nicotinamide adenine dinucleotide phosphate - reduced) is removed. When K = 4, one Glycolysis reaction (dhap → gap) that produces most portion of gap is removed. Both nadph and gap are important precursors in the biomass reaction. Removing these reactions causes the reduction of biomass flux values.

Table 2 summarizes the results from MOMAKnock. We first note that the MOMA flux distributions for all the suggested knockout strategies in fact have the corresponding succinate flux values that are consistently similar to objective function values in MOMAKnock without significant drops. Due to this, although the derived objective function values form OptKnock are higher, the final succinate productions for MOMAKnock suggested knockout strains under the MOMA objective are consistently better than OptKnock suggested knockouts except in the case K = 4, in which OptKnock derives an impractical strategy. The optimal succinate flux value from MOMAKnock suggested deletions can improve at least 37.5 percent compared to OptKnock in the MOMA flux distribution. In addition, both the succinate and biomass reaction flux values change smoothly for MOMAKnock strategies. Finally, as expected due to the L₂ distance based phenotypic constraints in the inner level of MOMAKnock, we can see that the optimal knock flux distributions from MOMAKnock is always closer to the wild-type flux distribution compared to OptKnock suggested knockouts.

Biologically, it is interesting to note that our MOMAKnock indeed identifies relevant reactions as suggested knockout reactions. For example, when the knockout number K is 2, one of the suggested knockout reactions is to eliminate the reaction that decompose the succinate (suc), and another one is to remove the reactions that involve competing byproduct metabolism for succinate such as 6-Phospho-D-gluconate (6pg) and Ribulose 5-phosphate (ru5p). With K = 3, MOMAKnock adds one additional knockout reaction to the previously identified ones based on the K = 2 case, which leads to the increase of succinate production to 32.15% of its theocratical maximum value. When K = 4, besides the reactions that consume succinate and the competing reactions, the reaction that decompose Phosphoenolpyruvate (pep) is also detected. This increases the succinate to 37.9% of the thearetical maximum. Finally, when K increases to 5, one more reaction is knocked out, which lead to 53.31 mmol/gDW · hr succinate produce rate. While as mentioned above, this reaction can convert biomass product to biomass precursor, so the deletion causes the reduction of the biomass flux rate.

Based on these preliminary results on this core network model, even though the OptKnock takes the maximizing biomass production as the inner cellular objective, the derived knockout strategies do not always achieve high biomass production when we simulate these knockout strategies under the MOMA objective. Sometimes, these knockout strategies cannot even guarantee the minimum biomass requirement. The reason for this is that the inner optimization in the bi-level framework of OptKnock serves as the additional constraint for the outer optimization problem. The derived optimization procedure first considers the outer problem as the primary objective and then the inner problem is optimized. The simulated low targeted chemical production rates for OptKnock suggested knockouts in the MOMA flux distribution and the abrupt biomass level changes in OptKnock illustrate that the biomass maximization assumption to approximate cellular objectives may not provide robust and reliable metabolic reaction deletion strategies. On the other hand, MOMAKnock approximates the inner cellular objective by the MOMA assumption which assumes that knockout strains stay closer to the corresponding wild-type strains. If this is guaranteed, knockout strains also can achieve appropriate biomass flux values. In fact, as shown in Tables 1 and 2, MOMAKnock not only achieves higher targeted succinate flux values under the MOMA objective but also obtains appropriate biomass flux values within the normal range compared to OptKnock. We also notice that with the increasing K, both the targeted succinate flux values and biomass values change smoothly contrasting to the abrupt changes in OptKnock, which may also serve as an evidence that MOMAKnock can help derive more robust knockout strategies with predictable performance.

By comparison with OptKnock on this core E. coli metabolic network, it is clear that our MOMAKnock may suggest more practical and robust knockout strategies for optimal bio-productions under phenotypic constraints.

Succinate production on iAKF1260 network

We further test MOMAKnock on a large E. coli metabolic network model--iAF1260 [19], which has 1,658 metabolicals and 2,936 reactions including the pseudo reactions required for the computation model. As in the core network model, succinate is set as the target chemical, the glucose uptake rate is fixed at 100 mmol/gDW · hr, and the minimum biomass is also set to 5 mmol/gDW · hr. All of our experiments are still based on the aerobic environment and all of the data are also from the OptKnock software package [13]. Table 3 provides the results from MOMAknockout for K = 3, 4, 5. Figure 2 shows the MOMA flux distribution for the wild-type strain as well as the MOMA flux distribution and the corresponding knockout reactions for the derived knockout strain with K = 5.

		MOMAKnock		MOMA Flux
K	Knockouts	Succi	Biomass	Succi	Biomass	${∥v-w∥}_{L_2}$
3	q8+succ→fum+q8h2, 6pgl+h2o→6pgc+h, (2)h2o + o2 + urate → alltn + co2 + h2o2	39.30	5.02	27.45	5.02	906.49
4	q8+succ→fum+q8h2, ac + atp → actp + adp, h2o+methf→10fthf+h, r5p+xu5p-D→g3p+s7p	67.08	5.02	63.23	5.02	402.33
5	q8+succ→fum+q8h2, glu-L+h→4abut+co2, 3pg+nad→3php+h+nadh, 3php+glu-L→akg+pser-L, 6pgc+nadp→co2+nadph+ ru5p-D	74.94	5.02	66.67	5.02	464.76

From Figure 2 and Table 3 we can see that, similar as in the core network model, MOMAKnock suggests knockout reactions in this large network that mostly contain the reactions that directly consume succinate, which include the succinate dehydrogenase reaction (SUCDi), as well as the competing reactions that may consume the precursors for succinate production, such as 6-phosphogluconolactonase (PGL), transketolase (TKT1) and phosphogluconate dehydrogenase (GND). The knockouts also contain some nonintuitive reactions as the final network dynamics is determined globally due to highly complex interactions among different reactions. When K = 5, the succinate production can achieve as high as 79.73% of the theocratical maximum rate (83.62mmol/gDW · hr), which demonstrates that our MOMAKnock can serve as a computational tool for deriving potentially effective and robust knockout strategies.

We notice that in Table 3 all of the biomass value is near 5 mmol/gDW · hr, which is the minimum biomass value set in the all of the tests. However, in this large network, the theoretical maximum biomass is 9.657mmol/gDW · hr. Experiments shows that if the the succinate dehydrogenase reaction (SUCDi) is recovered from knockouts, we can get higher biomass but the succinate can drop to as low as 10 mmol/gDW · hr. As shown in Figure 2, the reason for this is that the SUCDi reaction is the only direct pathway that can convert succinate back to some biomass precursors. Due to this reason, MOMAKnock derives the suggested knockout strategies, which try to find a point that can balance the succinate and biomass production.