Many limitations! Constructing the designs (i.e., going from data->dna->cells) takes time and is expensive: we pay vendors for the construction. So we have to restrict our software to output designs that yield the most novel products, with the least change to the biology of the microbe.

Most of the time changes do nothing! But if our models are accurate enough, the engineering can be predictable. We are continuously improving our models -- from more data mining, and from feedback from experiments.

That said, what our algorithm predicts right now is still a huge space (5000-10000 products). That is the tree image in the article. For context, the chemical industry centers around 70,000 products.