A Language Model Edits Designed Molecules Until Software Can Plan a Route

Ask a generative model for a molecule that will fit a particular protein and it will hand back thousands, each one written out as a tidy string of letters and brackets. Give that list to a medicinal chemist and much of it comes back untouched. The shapes are plausible, the predicted binding is attractive, and nobody can see where to start building them. The field's own phrase for this is synthetically inaccessible: real on a screen, and stranded there.

A study published on Sept. 23 in Nature Machine Intelligence goes after that shelf of stranded designs, and its answer is not a better designer. Junren Li and Luhua Lai of Peking University describe SynCraft, a system that takes a molecule somebody has already drawn and asks a large language model for a short list of changes to it: delete this atom, turn that one into nitrogen, close a bond here. Software then carries the changes out.

Everything rests on what counts as buildable here, and the definition is narrower than the word suggests. Success in this work is declared by a retrosynthesis program: software that starts at the finished molecule and works backward, one plausible reaction at a time, until it arrives at chemicals a laboratory can order from a supplier. Reach the catalog inside the time allowed and the molecule is scored as synthesizable. Run out of moves, or out of clock, and it is not. The authors' deposited dataset spells out the setup they used: every route is checked with their own planner, SimpRetro, against the Enamine catalog of building blocks, with a 30-minute budget per molecule on a single processor.

So the claim the paper earns is this: a program that could not find a route to a molecule can now find one, after a handful of atoms move. That is a genuine and useful thing. Just don't confuse it with a chemist actually synthesizing the stuff — nobody in this study did.

The authors call the effect they exploit the "synthesis cliff," a coinage of their own. It is built by analogy to the activity cliffs medicinal chemists already live with, where a trivial change to a molecule collapses its potency. The same steepness turns up in what can be built. One unusual ring, one awkward join between two fragments, and a planner working from libraries of known reactions has nothing that matches. Replace the ring with a common one and a route appears. The edit is tiny and the consequence is all or nothing.

That shape is why editing suits a language model better than writing does. Asked to invent a molecule outright, the model has to spell a valid chemical string character by character, which is the brittle symbolic work these systems are worst at. Asked instead for a sequence of edits, it has only to decide what to change and why; ordinary chemistry code applies the instructions, and the result is a real molecule by construction. Li and Lai write that predicting executable edit sequences "circumvents the syntactic fragility of LLMs while harnessing their chemical intuition."

On the benchmarks, the authors report that SynCraft "outperforms state-of-the-art baselines in generating synthesizable analogues with high structural fidelity," with "comparable performance across both proprietary and open-weight backends," meaning it works about as well on a commercial model as on one anybody can download. The yardstick is worth naming, though: SimpRetro, the planner that decides whether a molecule has become synthesizable, comes from the same laboratory, with the same first and senior author. Judging edited molecules by an in-house route finder is the normal protocol in this corner of the field, and it is also why the result belongs inside the computer until someone outside repeats it.

The more convincing of the paper's two demonstrations borrows its molecules from strangers. Its targets are high-scoring candidates designed to block a protein called RIPK1, which the molecular generation literature had discarded; they trace to a 2022 study in Nature

Communications by a group with no overlap with this one. SynCraft edited the discarded molecules until the planner found routes to them, while holding on to the parts that touch the protein. The paper calls this a "prospective" rescue — meaning it contrasts with a retrospective test, not that anything was actually put in a flask.

The second demonstration is closer to home. There SynCraft's edits to PLK1 inhibitors are set against the choices human medicinal chemists made in an earlier study called TransPharmer, and the authors report that the system "successfully replicates expert medicinal chemistry intuition." The senior author of SynCraft is also an author of TransPharmer, and the two papers are funded by the same grant, so the expert intuition being matched is in part the same laboratory's own.

There was no lab work behind this paper at all. No molecule was synthesized, none was assayed, and the whole evaluation is one program's verdict on another program's output. Checking it from outside is at least possible: the implementation is MIT-licensed on GitHub, and the test sets, the model's reasoning traces and a corpus of 3,332 synthesis-cliff pairs are posted on Figshare. Any other group can now run the same edits past a different planner, a different catalog of starting chemicals or a bench.