Do in a way so that it is as reuseable as possible (in the context of OBO), and that promotes clear communication (conveyance of meaning) and combination / merging of 'knowledge'.
We assume OBO Foundry principles for the conduct of the project.
Assume BFO as an upper ontology, not out of dogma but because they seem to work. When they stop working we consider something else.
Adopt the BFO attitude: individuals are particular things in the world (e.g. single molecules or particular populations of molecules). So proteins (in the AKT1 or whatever sense) are classes of molecules, not individual things. This has salutory benefits for subsumption, as you can do subclassing on classes but not subindividualing on individuals. (Think about alleles, isoforms, topoisomers, etc.)
Assume OWL, so that we can exploit semantic web technologies including SPARQL.
Interoperate with OBO ontologies, other things being equal. Interoperation is good. Specifically: RO, CheBi, OBI, EMBO, FMA, PATO, GO, EnvO