My reasoning goes like this:
In principle, any trapdoor function can serve as the foundation for an asymmetric encryption scheme.
All ciphers are trapdoor functions, since it's easier to encrypt plaintext than it is to crack ciphertext. Put another way, it's easier to use key and plaintext to produce ciphertext, than it is to go backwards and use ciphertext to produce key and plaintext.
Therefore, one could (at least in theory) devise a public key cryptosystem derived from any cipher, whether it be Caesar or AES.
This has the interesting implication that one could devise an asymmetric encryption scheme based on an existing asymmetric scheme + key + plaintext combo, and then derive another from that one, etc. etc.
I found this interesting quote which seems to confirm my hunch:
every problem X in NP has a naturally associated public-key cryptosystem (warning: it's a .PDF file)
H(key || counter) xor plaintext
is a stream cipher. You can use a four round feistel network to construct a block-cipher (see Luby-Rackoff), etc. $\endgroup$