Metabolic Pathway Design Using Active Subspaces

Salmonella can utilize a biodiesel waste product, glycerol, to produce 1,3-PDO, a common commercial solvent.  Experimental collaborators modify Salmonella to sequester the two enzymes needed for this pathway in microcompartments (MCPs), protein-bound shells that spatially segregate reactions. I formulated a differential equation model of this system to compare five candidate strains with different enzyme concentrations in the MCP.

To rank the strains, I considered quantities of interest (QoIs): product yield, rate of production and toxicity level. Evaluating QoIs on uniformly sampled parameters, restricted by physical constraints and prior measurements, is computationally intractable. To efficiently generate value ranges for each QoI, I used Active Subspaces to identify parameter directions that most affect each QoI. I then used maximin sampling to produce a space-filling spread of parameter samples in the significant parameter directions while randomly perturbing the samples in the insignificant directions. This sampling reduced the computational load by at most 5 orders of magnitude when compared to a coarse grid sampling. The QoI distributions converged with increasing numbers of samples. I used hypothesis testing on the QoI distributions to predict optimal producing strains.