I sat in on a review where an ML team presented a diffusion model generating candidate small molecules for a target binding site — genuinely impressive output, chemically valid structures, good predicted binding affinity, novel enough to be patentable. Then a medicinal chemist on the call asked, politely, how many of these anyone could actually synthesize in under six months. The room got quieter than I expected.

Valid isn't the same as makeable

The generative model's objective was, reasonably, to produce chemically valid structures with strong predicted binding properties. Nothing in that objective accounts for synthetic accessibility — whether a route to actually build the molecule in a lab exists using reasonably available starting materials and a reasonable number of steps. A structure can be perfectly valid on paper and still require a synthesis route nobody has ever demonstrated, built from a strained ring system that decomposes readily, or need a chiral center installed with a reaction that only works in theory.

The teams doing this well now bake a synthesizability score into the generation or filtering process directly — something like a retrosynthesis-model-derived score, not just a post-hoc filter applied to the output. Bolting a synthesizability check onto the end of a generation pipeline that wasn't optimized for it produces a lot of near-misses: molecules that fail the check by one step, one reagent, one unstable intermediate.

Where this is genuinely working today

  • Scaffold hopping — starting from a known, synthesizable core and generating variations, rather than generating from nothing. Much higher hit rate on synthesizability because you're not straying far from known chemistry.
  • Fragment-based generation constrained to a library of known, purchasable building blocks, rather than free-form atom-by-atom generation.
  • Using the generative model to produce a large candidate pool, then leaning hard on retrosynthesis models to filter down to the subset worth a chemist's time, instead of expecting a chemist to manually triage hundreds of candidates.

Fully free-form, atom-by-atom generation with no synthesizability constraint makes for a better demo than a better pipeline, at least with where the field is right now. The molecules are real and often genuinely novel — the gap is entirely in the step between "this is a valid molecule with good predicted properties" and "this is a molecule a lab can actually make and test," and that gap is where I'd focus if I were building in this space today rather than on generation quality itself.