Drug-design tools are sharper than the scores judging them

Peptides, with the reasoning attached

In a preprint, researchers describe CIPER, a model for optimising cyclic peptides that predicts how a specific structural edit changes permeability. The point is not only the prediction. The model is built to show its working, using atomic and cyclic-residue kernels to give signed, interpretable contributions for each edit, so a chemist can see which change is supposed to help and why. m:…2542

The reported gains are large: a mean permeability improvement of 0.518 log10 units per query in cyclic peptide libraries, and a 75.1% reduction in regret compared with random selection over four queries. That is the kind of result that changes which candidates get synthesised first, and the interpretability is what makes it usable rather than merely impressive, since a team can prioritise edits it understands. m:…2542

The contrast with a black-box ranking is the story here. A model that explains its choices lets a team catch a plausible-looking mistake before it reaches the bench, while a model that only scores cannot. m:…2542

One ratio, two answers

A second preprint posted the same day shows what happens when the score itself is ambiguous. Computational models of P-glycoprotein efflux are often validated against the efflux ratio from bidirectional monolayer assays, but the study finds that ratio supports two distinct mathematical readings: net flux asymmetry and the pump term, related by the saturating function f = x/(x + 2). Validation against the ratio is ambiguous unless the reading is stated explicitly. m:…2541

The two readings do not differ by a rounding error. Across the compound pairs studied, the pump-term reading yielded apparent energy differences up to 3.41 kcal/mol, while the net-asymmetry reading remained below 0.43 kcal/mol for the same pairs. A model validated under one reading can look strong and under the other look weak, with nothing about the model itself having changed. m:…2541

The fix is unglamorous: state which reading was used. But the implication is broad. Any team comparing efflux models, or buying one, should ask for the validation reading before trusting the headline number, because the number alone does not say which question was answered. m:…2541

Can it even be made?

A third preprint addresses a gap further downstream: whether a designed molecule can be synthesised at all. Synthmc is a precedent-based synthesizability score for transition metal complexes, drawing on the Cambridge Structural Database. Traditional organic scores are insufficient for these complexes, the authors argue, because they ignore metal centres, coordination modes and steric crowding. The score assigns lower values to more recently reported complexes, which the authors read as an ability to identify novel chemistry. m:…8150

A fourth preprint, AutoSPS-D2B, closes the loop toward the bench. It integrates automated solid-phase synthesis into direct-to-biology drug discovery, enabling multistep, resin-bound chemistry and direct assaying of crude libraries. The workflow sampled over 1,700 building block combinations, a scale of sampling that solution-phase methods in direct-to-biology screening do not reach. m:…8149

Applied to CCR2 ligands, the platform improved binding affinity six-fold within two DMTA cycles. That is a single target class, not a general guarantee, but it shows what automated synthesis adds when the chemistry it enables is genuinely broader rather than merely faster. m:…8149

What this changes

Read together, the four preprints describe a design pipeline that is getting more honest about its own limits. Prefer models that explain their choices. Ask which validation reading a score used. Filter candidates for synthesizability with a score built for their chemistry, not a borrowed organic one. And automate the synthesis step only where it genuinely widens the reactions available. m:…2542 m:…2541 m:…8150

All four are preprints, not peer-reviewed papers. The peptide gains come from library optimisation, the efflux finding is a reanalysis of assay interpretation, and the six-fold affinity gain is from one target class. The pattern is worth acting on, but the numbers are early evidence, not settled performance. m:…2542 m:…8149