Abstract

Motivation: Random sampling of the solution space has emerged as a popular tool to explore and infer properties of large metabolic networks. However, conventional sampling approaches commonly used do not eliminate thermodynamically unfeasible loops. Results: In order to overcome this limitation, we developed an efficient sampling algorithm called loopless Artificially Centered Hit-and-Run on a Box (II-ACHRB). This algorithm is inspired by the Hit-and-Run on a Box algorithm for uniform sampling from general regions, but employs the directions of choice approach of Artificially Centered Hit-and-Run. A novel strategy for generating feasible warmup points improved both sampling efficiency and mixing. II-ACHRB shows overall better performance than current strategies to generate feasible flux samples across several models. Furthermore, we demonstrate that a failure to eliminate unfeasible loops greatly affects sample statistics, in particular the correlation structure. Finally, we discuss recommendations for the interpretation of sampling results and possible algorithmic improvements. Availability and implementation: Source code for MATLAB and OCTAVE including examples are freely available for download at http://www.aibn.uq.edu.au/cssb-resources under Software. Optimization runs can use Gurobi Optimizer (by default if available) or GLPK (included with the algorithm).