While it's easily possible to load and save in dict form and convert dicts to pandas DataFrames, it is quite slow, by bypassing BSON > dict conversion and going straight from BSON > DataFrame there is great memory and performance benefits.

I have seen this guide on how to load data to pandas although this has 2 caveats, it isn't native, and there isn't an alternative for saving using pandas. 

There is a default_class argument for MongoClient, although this is restricted to MappedCollection.

I've also seen Monary but this has no support for string types and while not under active development for 4+ years I'm unsure the compatibility with the newest version of Pymongo.