An LLM that edits molecules instead of drawing them
Nature Machine Intelligence has published a Peking University method that reframes molecular design as editing rather than generation — and the numbers suggest the framing is the contribution. SynCraft, from Junren Li and Luhua Lai, takes a molecule that scores well but can't be made and asks a language model for the shortest set of structural repairs that make it synthesizable: delete an atom, mutate one, add a bond, change a bond order, set a chiral centre. The model never writes a structure string. It writes a short JSON sequence of edit commands, and RDKit applies them deterministically, so every candidate the model proposes is a valid molecule by construction and the retrosynthesis check that follows is real chemistry rather than text matching.
The design choice worth noticing is what the model does not control. Generation stays cheap and unconstrained; verification is where the work happens. In their setup, Gemini-2.5-Pro reasons over the molecule's specific liabilities — reactive groups, dead-end fragments, hidden synthetic cliffs — then emits the edit sequence, with five worked examples retrieved at run time to steer the format. The paper reports the model performing comparably across proprietary and open-weight backends, though that open-weight comparison sits behind the journal paywall and is absent from the preprint.
The results land where the field's bottleneck has been. Against ChemProjector, SynFormer and ReaSyn, SynCraft finds a synthesizable analogue at a Tanimoto similarity above 0.5 for 42.7% of the AI-designed molecules in the Pocket2Mol set, where SynFormer manages 30.7%; on the ResGen set it reaches 44.7% against ReaSyn's 37.5%. At the looser similarity threshold of 0.6 the gap widens to 28.4% versus 14.1%. One honest asterisk: at the strictest threshold the older baseline wins on one dataset, 4.3% to 3.5%, so the advantage is about the bulk of candidates rather than the closest analogues. Training used 3,332 paired examples, and the test sets are just two of five molecule generators — the other three produced too few synthesizable molecules to test on.
What makes the case studies convincing is the audit trail. The model reproduced a medicinal chemist's own repair to a PLK1 inhibitor step for step, and when given interaction-aware prompting it recovered 42 shelved, high-scoring RIPK1 candidates whose docked binding survived the edit. There is also an end-to-end run on SARS-CoV-2's main protease. All of it is computational — no molecule here was made in a flask — so for now SynCraft is a better filter in front of the wet lab, not a replacement for it.
What to watch: whether design-in-the-loop pipelines adopt a deterministic edit layer as a standard guard rail, and whether anyone publishes an open-weight run that reproduces the journal's cross-backend claim.
If an AI-designed drug only reaches the clinic after a second model rebuilt it, does the first model deserve credit for the discovery? Tell us in the comments.
Sources: SynCraft-Core code (GitHub)
