Biochemical parameter estimation vs. benchmark functions: A comparative study of optimization performance and representation design

Computational Intelligence methods, which include Evolutionary Computation and Swarm Intelligence, can efficiently and effectively identify optimal solutions to complex optimization problems by exploiting the cooperative and competitive interplay among their individuals. The exploration and exploitation capabilities of these meta-heuristics are typically assessed by considering well-known suites of benchmark functions, specifically designed for numerical global optimization purposes. However, their performances could drastically change in the case of real-world optimization problems. In this paper, we investigate this issue by considering the Parameter Estimation (PE) of biochemical systems, a common computational problem in the field of Systems Biology. In order to evaluate the effectiveness of various meta-heuristics in solving the PE problem, we compare their performance by considering a set of benchmark functions and a set of synthetic biochemical models characterized by a search space with an increasing number of dimensions. Our results show that some state-of-the-art optimization methods – able to largely outperform the other meta-heuristics on benchmark functions – are characterized by considerably poor performances when applied to the PE problem. We also show that a limiting factor of these optimization methods concerns the representation of the solutions: indeed, by means of a simple semantic transformation, it is possible to turn these algorithms into competitive alternatives. We corroborate this finding by performing the PE of a model of metabolic pathways in red blood cells. Overall, in this work we state that classic benchmark functions cannot be fully representative of all the features that make real-world optimization problems hard to solve. This is the case, in particular, of the PE of biochemical systems. We also show that optimization problems must be carefully analyzed to select an appropriate representation, in order to actually obtain the performance promised by benchmark results