a lot of #MachineLearning & #ML4Molecules work is based on the idea that molecular structure encoded in eg SMILES is the essence of chemistry.

just FYI for CS people, graphs are to molecules what Turing machines are to computers

a convenient and highly idealized abstraction https://twitter.com/leecronin/status/1338289820746059778

firstly: not all molecules are representable by SMILES (non integer bonding orders, metal bonding, etc)

secondly: not all molecules are representable by 2D graphs (conformational isomers / atropisomers, non-covalent complexes, delocalized ionic and radical species, etc)

thirdly: not all molecule properties can be derived from static 3D structures (crystals, solvent cages, entropic effects, excited state properties)

my point is that it goes beyond just the experimental part that @leecronin describe : common representations *can't even* theoretically describe many important properties in #compchem , let alone experimental ones

(there are many very good ML papers that do take these things into account BTW from the likes of @adrian_roitberg, the @ZimmermanUMich group, the R Gomez-Bombarelli and HJ Kulik groups, etc.)