SynCraft edits molecules atom by atom instead of asking an LLM to write new SMILES

A language model asked to write a SMILES string has to get every bracket, ring-closure digit and valence right in a single pass, and one bad token produces something a cheminformatics toolkit will refuse to parse. SynCraft asks for edits. The model predicts a sequence of executable atom-level operations applied to an existing molecule, and the toolkit performs the chemistry ^[2]. The authors wrote that the framework "circumvents the syntactic fragility of LLMs while harnessing their chemical intuition" ^[3].

The thing being hunted is what they call the synthesis cliff, where minimal structural modifications yield substantial gains in synthetic feasibility ^[6]. They are making a claim about the shape of the scoring surface. The claim also invites a sceptical reading of the benchmark, because a feasibility score that jumps on a one-atom swap is sensitive to features that may or may not correspond to an available route. The abstract gives no benchmark figures, no name for the synthesizability metric behind them, and no compound made in a flask ^[14][15].

SynCraft is aimed at a specific complaint. A substantial fraction of generated molecules is synthetically inaccessible ^[4], and the standard remedies, post hoc filtering and projection onto predefined synthetic templates, compromise structural novelty or disrupt key pharmacophores ^[5]. Editing a parent locally is a plausible answer to that, and the supporting evidence released with the paper includes PLIP retention tables ^[11]. Those tables report interaction patterns computed from structures, and it would take an assay to report binding.

Comparable performance across proprietary and open-weight backends ^[8] is the finding with the most immediate operational consequence, since a group with unpublished series can then run the edit loop on machines it controls. The released training corpus is 3,332 synthesis-cliff pairs with reasoning traces ^[11]. For chemistry that is a small corpus, and I would expect most of the chemical knowledge to come from the backend model's pretraining, with the pairs teaching the edit format.

The case studies push the claim furthest. Using interaction-aware prompting, the authors report replicating expert medicinal chemistry intuition in editing PLK1 inhibitors and rescuing high-scoring RIPK1 candidates that had previously been discarded in the molecular generation literature ^[9]. Taken at face value, the reject pile from earlier generative campaigns becomes somewhere to look for candidates. The same pipeline, packaged as an agent skill, was run end to end on molecules generated for the SARS-CoV-2 main protease ^[10].

The implementation, the reproduction scripts for every experiment and the agentic skill package are on GitHub under an MIT licence, with a Zenodo archive ^[12]. The paper itself, on nature.com, sells for $39.95 ^[13].