Teaching a language model to carry potency across scaffolds

Changing a molecule’s central scaffold poses a practical question: which lessons from the old series should survive? A substituent that improves potency in one family is a hypothesis when moved into another. Grebner and colleagues at Sanofi address that problem with a language model, then put selected designs through synthesis and biochemical assays. Their study appeared in Communications Chemistry on 15 September 2026 as a peer-reviewed Article in Press.¹

The work deserves attention because it connects a conversational design method to compounds that were actually made. Its strongest contribution is a useful way to bring existing structure–activity relationships into a new chemical series. The measured results support the combined design-and-selection workflow; they do not isolate the language model from the chemists, predictive models and structural calculations that helped choose the compounds.¹

What the model learns from the conversation

The method uses Claude 3.5 Sonnet with molecular structures written as SMILES strings. In-context learning supplies examples during a query while leaving the model’s parameters unchanged.¹ The authors’ public implementation also separates molecular generation from predictive models built on fingerprints, molecular descriptors and learned fragment representations.²

That separation matters to a medicinal chemist. A model can suggest an analogue without being the most accurate predictor of its potency. Generating a useful idea, ranking a set of ideas and explaining an observed structure–activity relationship are different jobs. Evaluating all three under a single label such as “chemical intelligence” would hide the decisions that make a design process useful.

There is a clear methodological predecessor. In a 2024 preprint, Moayedpour and colleagues explored many-shot molecular design using examples supplemented by generated molecules carrying predicted activities. Different predictive models screened which proposals entered the next round. That work assessed computational design and prediction, including comparisons with REINVENT 4; its generated-molecule results were predictions rather than prospective biochemical measurements.³

REINVENT 4 provides a different established route to guided generation: it combines recurrent or transformer generators with transfer learning, reinforcement learning and staged optimization. Its architecture makes scoring functions and molecular constraints explicit components of design.⁴ The interesting comparison is therefore about how project knowledge enters the process. A conversational interface can make a chemist’s instructions easier to express, but it still needs a disciplined relationship between those instructions, the scoring machinery and the experiment.

Asking for the right chemical neighbourhood

For cathepsin A, a serine protease, the new study draws on an established biaryl series to improve a pyrazolone series. The language model receives 296 biaryl examples and ten weakly active pyrazolone starting compounds. A larger project dataset of 2,465 compounds supports separate predictive models; it is not all supplied to the language model.¹

This is a well-chosen medicinal chemistry problem. The desired answer is an analogue of the new series that benefits from knowledge elsewhere in the project. Simply returning another member of the established family would miss the point, even if that molecule looked convincingly potent.

The prompt comparisons expose that distinction. Without explicit scaffold rules, generated structures gravitated towards the example series. Adding structural requirements redirected the proposals towards the intended pyrazolone chemistry, at the cost of more repeated structures. Supplementary analyses also test how rules about fluorine and carboxylic groups alter the output.¹⁵

That trade-off should be judged against the assignment. Diversity is valuable when a team wants alternative starting points; a focused set can be more useful when chemistry resources are committed to a tractable series. Neither a larger collection nor a narrower collection is intrinsically better. The practical question is whether the generated set gives the chemist meaningful choices within the part of chemistry they can investigate.

Rules also need checking after generation. A natural-language instruction is a request, not a structural validator. The study’s use of canonicalization, deduplication and downstream selection is consequently part of the method’s credibility, rather than an embarrassing correction to an otherwise autonomous process.¹ It gives the conversational step a defined role inside an executable chemistry workflow.

Where predicted feedback meets the assay

The cathepsin A procedure iterates computationally. Promising generated structures return to the candidate pool with predicted activities; new assay measurements do not arrive between those iterations. Separate predictive models remain fixed and participate in the final selection.¹⁵

This distinction changes how the apparent progress should be read. An improving predicted score is evidence that the search is following its computational objective. It becomes evidence of improved inhibition only when the proposed compound reaches the assay. Keeping the final predictors anchored to experimental project data is a sensible safeguard against simply retraining every evaluator on the generator’s own guesses, although agreement between predictors cannot substitute for measurement.

The selected prospective set gives the method substance. Ten newly synthesized cathepsin A designs were tested, six with IC50 values below 1 µM; compound 11 reached 162 nM. Nine of the ten improved on their corresponding starting compounds. Twelve additional designs had already existed in historical project records and were absent from the supplied training data; those rediscoveries are useful corroboration but are a separate group from the newly made compounds.¹⁵

There is also a chemically relevant comparison. Using the selected building blocks, the team made 27 additional combinations, of which six were submicromolar. Thus the selected designs yielded six successes from ten compounds, versus six from 27 in the combinatorial expansion.¹ The comparison asks a practical question: did the chosen combinations concentrate useful activity within accessible chemistry?

The result is encouraging, and the scope of the control matters. Both sets share building blocks already selected through the design process. This is not an independent contest against an unconstrained medicinal chemistry team, and the study does not measure a project-wide saving in time or synthesis effort. The defensible conclusion is enrichment within this selected chemical exercise. That is already a worthwhile contribution: making fewer weak combinations can be valuable without establishing that every stage of discovery has become faster.

The substitutions, and the explanations attached to them

The supplementary examples make the chemistry more tangible. For compounds 11 and 12, the proposed changes include replacing a bulky tert-butoxy group with methyl and introducing fluorine on an aromatic ring. Another example exchanges a cyclopentyl-containing part for an aromatic alternative. These are recognizable medicinal chemistry proposals, tied to the supplied examples rather than presented as mysterious discoveries.⁵

A useful explanation should help a chemist decide what to test next. It need not be a complete mechanism. Here, an explanation that associates fluorination with a favourable project trend can point towards a sensible comparison. Claims about improved metabolic stability or a particular binding interaction remain hypotheses unless the corresponding property or interaction is measured. Potency alone cannot tell which of several simultaneous molecular changes caused an improvement.

The authors also show an unsuccessful explanation. For compound 20, the model favours replacing tert-butoxy with tert-butyl, but its predicted activity exceeds the experimental result. The supplementary discussion explicitly identifies the suggestion as wrong.⁵ Including this example improves the paper: readers can see how a plausible narrative may be useful for inspection without being a reliable account of what the molecule will do.

There is one specific sample-quality qualification worth retaining. The supplementary characterization flags compound 19 as low purity, with a 5.6:1 mixture involving the tert-butoxy-containing material and a phenolic component; comparator compounds 37 and 44 have related purity qualifications.⁵ Their assay values should therefore be treated as results for the tested samples, not clean attributions to single pure structures. Purification and retesting would sharpen those local SAR inferences. This qualification does not erase the separate result for compound 11.

A second target, with a different outcome in each series

The RIPK1 experiment broadens the study beyond one protease series. The researchers used a context dataset of 2,552 project compounds, excluding the requested scaffold families and closely related motifs. They synthesized 16 selected designs in each of two series. The pyrrolidine set included nine compounds below 1 µM, with a best IC50 of 9 nM. The best imidazolone compound was 4.29 µM.¹

Keeping those outcomes separate is more informative than pooling them into a success rate. A productive design method need not perform identically in different chemical families. The pyrrolidine result gives a strong reason to continue that chemistry; the more modest imidazolone result sets a different starting point for the next decision. The contrast is valuable evidence about where the workflow found useful compounds, rather than an obligation to manufacture a negative verdict.

The selection process is especially important here. Generated proposals passed chemical filters, docking, energy-based scoring, visual inspection and consideration of available building blocks and convergent routes. Predicted permeability and lipophilicity-related properties also informed prioritization.¹ The measured hits belong to this combined process. Describing them as the unassisted output of a chat prompt would omit much of the medicinal chemistry that made the experiment work.

Likewise, a computed pose can help prioritize a proposal without establishing its binding mode experimentally. The useful interpretation is that ligand examples and structural modelling contributed complementary information. The study supports putting those tools together; it does not require choosing a winner between them.

What a potency result contributes to the next decision

The assays measure biochemical inhibition. RIPK1 was assessed through an ADP-Glo assay with 50 µM ATP; cathepsin A used a fluorescence assay at pH 5.5 with 2.5 µM substrate. Measurements were performed in duplicate, and the cathepsin A designs are described as stereoisomeric mixtures.¹ These conditions define the results and their comparisons. The best numbers from the two targets should not be arranged into a universal ranking of molecular quality.

For the next medicinal chemistry cycle, the appropriate question is what to preserve while refining the series. Potency provides a reason to invest further; stereochemistry, selectivity and measured disposition properties determine which changes remain useful as the project progresses. The paper itself places broader optimization, including solubility and metabolic stability, in future work.¹ That is a boundary of this investigation, not a defect in an experiment designed to test prospective molecular design.

The public repository makes the workflow more inspectable by identifying the model configuration and organizing generation, prediction and analysis code.² The reporting summary identifies the project SAR data as confidential, so the public release does not expose every input to the prospective work.⁶ Availability does not mean that this review reproduced the computations or laboratory results. The assessment rests on the full paper, its relevant supplementary methods and characterization, and the earlier methodological work.

My reading is positive. The study gives chemists a concrete way to specify which scaffold should remain, which examples should inform it and which proposals merit experimental attention. Its value lies in carrying project knowledge into a new set of testable molecules. The strongest lesson is a practical one: conversational design earns a place in medicinal chemistry when it improves the choices that reach the bench and leaves the resulting measurements available to challenge its next suggestion.