How does MP predict metastable structures that have no natural analogue?

Hi MP folks,

How does MP predict the metastable crystalline phases that have never been experimentally observed but are listed in the database? Are they simply based on substitutions from crystal structures in similar material systems, or are they predicted from first-principles ab-initio DFT?

I am asking because I would like to learn to do such ab-initio DFT predictions of novel metastable materials, and am wondering if pymatgen can help with that. I have played around a fair amount with the crystal toolkit and can definitely run my own DFT geometry / energy calculations on the hypothetical structures I can create, but that is a very slow process when considering all the possible metastable structures that are out there for multinary compounds.