Is OptKnock a reliable strategy for desirable mutants?

Flux balance analysis (FBA) has enabled the development of computational methods for predicting optimal knockout strategies to genetically engineer microbial strains for desirable behavior, such as optimal biochemical overproduction for alternative energy sources. Many of these existing methods are based on bi-level optimization formulations to maximize the desired biochemical overproduction at the outer-level while modeling cell survival as the inner-level optimization problem, for example, by maximizing cell growth as in the seminal OptKnock. Nevertheless, optimal knockout strategies derived in such a bi-level optimization framework may be heavily depending on the closeness and robustness of the inner-level optimization model in capturing actual cell survival states. We investigate how reliable the knockout strategies derived by OptKnock are, considering two critical but overlooked factors: (i) the surviving mutant of the inner-level optimization model may not be well-defined, i.e., it is not unique; and (ii) we cannot guarantee that the nature always cooperates with the human desire to select the microbial strains that produce maximum biochemical products among surviving mutants. We present our study in a core E. coli metabolic network and show that the knockout solutions from OptKnock could be of arbitrarily poor performance. Then, we revamp OptKnock through a novel pessimistic bi-level optimization framework, which considers the non-unique and noncooperative issues and potential modeling errors. Through computing pessimistic knockout solutions and benchmarking with those from OptKnock, we observe that they are more reliable and perform significantly better. We believe that the proposed pessimistic bi-level optimization framework will help identify more practical and robust knockout strategies.