Robust flux balance analysis of multiscale biochemical reaction networks

Yuekai Sun, Ronan M. T. Fleming, Ines Thiele, Michael A. SaundersBMC Bioinformatics 2013 · DOI

TL;DR

This paper introduces "lifting techniques" to make complex, multi-scale biochemical network models solvable by standard optimization software. By reformulating poorly scaled constraints, they enable accurate flux balance analysis (FBA) and flux variability analysis (FVA) for integrated models of metabolism and macromolecular synthesis, which were previously intractable.

What Problem Does This Paper Address?

Biological systems are incredibly complex, involving networks of reactions that operate at vastly different scales. Modeling these systems, especially when integrating different biological processes like metabolism and macromolecular synthesis, leads to "multiscale" mathematical models. These models are represented by large matrices of stoichiometric coefficients, where some values can be millions of times larger or smaller than others. This extreme difference in scale, known as "poor scaling," causes significant problems for standard optimization algorithms, which are the workhorses for analyzing these models (e.g., Flux Balance Analysis - FBA). Even state-of-the-art solvers, when applied to these poorly scaled models, can produce inaccurate or completely wrong results, reporting infeasible solutions when they are actually feasible, or vice-versa. This limits our ability to accurately predict the behavior of complex biological systems.

Methods

The core innovation presented is a set of "lifting techniques." These techniques aim to transform the poorly scaled constraints in the linear programming problem that defines FBA into a more manageable form. Instead of directly scaling the existing coefficients (which can sometimes worsen the problem for other parts of the model), lifting techniques introduce auxiliary variables to create a higher-dimensional representation of the feasible solution space. This effectively "lifts" the problem into a new space where the constraints are better behaved.

Specifically, they address two main types of poorly scaled constraints:

  • Mass balance constraints: For reactions with very large stoichiometric coefficients (e.g., A + 10000B → C + D), they decompose these into a series of reactions involving "dummy" metabolites. This breaks down the large coefficients into smaller, more manageable ones.
  • Coupling constraints: These constraints enforce consistency between different biological processes (e.g., between a metabolic reaction and the synthesis of its corresponding enzyme). When these involve ratios of fluxes that vary over many orders of magnitude (e.g., 0.0001 ≤ v1/v2 ≤ 10000), they are reformulated using auxiliary variables to create a sequence of constraints with more reasonable coefficients.

The authors implemented these techniques within the openCOBRA toolbox and ensured that standard optimization solvers (CPLEX and Gurobi) had their automatic scaling and presolve options disabled to work with the reformulated problem.

Datasets

The paper evaluates their methods on two integrated genome-scale models of E. coli:

  • ME76664: A model with 76,664 reactions.
  • ME76589: A model with 76,589 reactions.

These models represent the function of nearly 2,000 E. coli genes and include approximately 60,000 "coupling constraints" to ensure consistency between metabolism and macromolecular synthesis. These models are notable for their multiscale nature, with reaction rates varying over four orders of magnitude, and for containing a large number of "large matrix entries" (over 41,000 entries exceeding a threshold of 1024).

Key Insights and Ideas

  • Lifting as a general solution for multiscale optimization: The paper demonstrates that lifting techniques can effectively address poor scaling in optimization problems beyond just FBA, making them broadly applicable to complex biological models.
  • Preserving sparsity: While lifting adds variables and constraints, the new constraints are designed to be very sparse, minimizing the impact on the overall computational burden for large-scale solvers.
  • Beyond scaling: Traditional scaling methods can sometimes improve one aspect of a problem while worsening another. Lifting offers a more robust approach by fundamentally reformulating the problematic constraints.
  • Integrated models are the future, and they are hard: The increasing complexity of biological models (integrating metabolism, signaling, macromolecular synthesis, etc.) will only exacerbate scaling issues, making robust analysis techniques like these essential.

Key Results

The lifting techniques significantly improved the performance of standard optimization solvers (CPLEX and Gurobi) on the multiscale E. coli models:

  • Reduced infeasibilities: For FBA, the lifting techniques dramatically reduced or eliminated infeasible reports and significantly decreased the infeasibility of computed steady states. For example, in Table 1, infeasibilities dropped from 1.3 x 10⁻⁴ to 2.9 x 10⁻⁶ for the simplex solver.
  • Stabilized iterations: For FVA, the lifting techniques stabilized the number of simplex iterations and reduced infeasibilities. For instance, in Table 3, infeasibilities for a specific flux objective dropped from 1.1 x 10⁻⁴ to 7.6 x 10⁻⁵.
  • Improved solver reliability: While the barrier solver's performance was more varied, the simplex solver showed consistent improvements in accuracy and reliability after lifting.
  • Shrinking solution space: Figure 2 illustrates that integrating macromolecular synthesis into metabolic models substantially shrinks the feasible steady-state solution space, leading to more constrained and biologically realistic predictions.

Limitations and Future Work

  • Increased problem size: Lifting techniques do increase the number of variables and constraints, although the added constraints are sparse. This could still be a factor for extremely large models.
  • Barrier solver performance: The paper notes that the CPLEX barrier solver's performance was more varied and less reliable on the lifted problems compared to the simplex solver. This suggests that solver-specific behavior can still be an issue.
  • Future work: The authors suggest that their techniques could be combined with other advanced methods like iterative refinement or specialized solvers (e.g., those using rational arithmetic) for even greater accuracy, though they highlight the potential performance drawbacks of the latter. They also emphasize the growing need for such robust analysis methods as biological models become more comprehensive.

Summary updated

BibTeX
@article{sun2013robust,
  author = {Yuekai Sun and Ronan M. T. Fleming and Ines Thiele and Michael A. Saunders},
  title = {{Robust flux balance analysis of multiscale biochemical reaction networks}},
  journal = {BMC Bioinformatics},
  year = {2013},
  doi = {10.1186/1471-2105-14-240},
}

Loading related content...

Figure