>Won't some candidates simply be too complex and unstable to reliably produce using existing mammalian or baculovirus platforms?

You can add that to the objecting function that your model training function is optimizing, ensuring model output is not "too complex or unstable to reliably produce".