The choices that make a model useful
Chemical judgment enters a model before the result appears. This issue follows the choices that make that result worth using.
The archive.
Research worth revisiting, with its evidence intact.
A better optimizer needs a trustworthy test
RECOB puts algorithm rankings through two different checks—and shows why a stable virtual winner is only part of the evidence for a laboratory choice.
Letting a molecular generator choose its own size
A compact latent model generates molecules of varying size. Its careful separation of reconstruction, refinement and quantum-verified selection shows where the useful candidates come from.
Teaching a force field where water meets metal
SoLiD26 improves predicted atomic forces at solid-liquid interfaces. Its larger-cell test, curation choices and raw-force handling explain what that improvement establishes.
A molecular glue earns its shape in the cell
VAV1 degraders connect a predicted protein interface to chemical changes and cellular measurements—and show why better degradation can change the selectivity question.
What a receptor-state model learns about a drug
DSQ improves agonist bioactivity prediction by giving its neural network two receptor states. Its assays, ablations and transfer tests explain what that advance means.
The comparison is part of the discovery
Good controls deserve some of the attention usually reserved for ingenious models. Mira introduces an issue about evidence that helps a scientist decide.
What a formulation robot can learn from yesterday’s experiments
An evidence-ablation experiment makes a persuasive case for laboratory memory—while matched campaigns reveal why the best sample is only part of the result.
What a reversed sequence can teach a molecular model
An unreviewed study finds that structured practice before chemical training can improve property prediction. Its controls show why the benefit depends on more than extra training.
Teaching a phase map to respond to pressure
Exponentially tilted thermodynamic maps make environmental controls part of a generative model. Their practical promise becomes clearer when the reference simulation, sampled observable and computing budget stay in view.
What counts as progress when small changes already work?
A measured survey of close analogues gives AI-assisted medicinal chemistry a tougher comparator: accessible chemistry can improve potency while exposure properties move the other way.
What an organ-targeting model actually predicts
An interpretable lipid-nanoparticle classifier offers useful formulation hypotheses. Its outcome label and validation design determine how far they travel.
What makes a chemistry lab self-driving?
A new optimizer sharpens experimental choices. Two robot studies show why choices are only part of autonomy.
Generating a molecule is the beginning
SmileyLlama makes molecular generation steerable in words. The next question is how a proposed structure earns experimental trust.
A force field needs the right exam
Three materials studies ask whether a convincing simulation answers the physical question that matters.
Uncertainty that follows the atomic environment
A calibration method makes force-error estimates more responsive to local structure.
Six ions move together in a lithium-fluoride simulation
A predicted transport mechanism offers a question for battery-interface experiments.
Learning an equation for copper
Symbolic regression makes the form of an interatomic model part of the search.
From a literature procedure to a robotic test
A chemputation workflow tests the journey from extraction to physical execution.
A mass spectrum still leaves a choice
DiffMS generates molecular candidates from a spectrum and a known formula. Its benchmark gains are useful partly because the remaining ambiguity is measurable.
Adding orbitals to the molecular graph
Stereoelectronic features can enrich a molecular representation. A learned surrogate makes the extra information accessible while introducing another prediction boundary.
Choosing a plate, not just an experiment
Minerva makes parallel reaction optimisation a problem of batches and competing objectives. The analyst still has to define what success means.
Few-shot selectivity still needs a history
A meta-learning study asks how prior reaction data can help classify asymmetric hydrogenations when a new task has few examples.
Finding an enzyme for the substrate in hand
CATNIP links measured reaction compatibility to chemical and protein similarity. Its useful output is a ranked starting point for testing.
The test set is part of the question
DataSAIL makes molecular similarity an explicit part of evaluation. Its addendum also shows why no split should be treated as a universal answer.
What a bag of atoms can teach a benchmark
A 2024 study asks how much molecular topology a benchmark really demands, and tests a set-based alternative to familiar graph models.
When a highlighted atom becomes a hypothesis
An arene-selectivity model offers clues about where chemistry happens. A clue becomes more useful when a separate calculation can challenge it.
Zero-shot generation is not data-free
SiMGen generates molecular structures through local similarity. Its use of pretrained components makes the meaning of zero-shot worth spelling out.
Count the compounds before the hit rate
Nine synthesized A2A ligands make a useful case for reporting the choices between a generated library and an experimental claim.
An activity cliff is not a direction of travel
A 2026 preprint separates finding a sensitive molecular position from predicting a beneficial substitution.
A macrocycle can fit before it can travel
RFpeptides provides measured binders and structural confirmation. Permeability remains a separate design question.
A potency gain needs a chemical history
A MAGL optimization study joins reaction prediction to compound design. Its most useful record is the route from a proposed substitution to a measured series.
The selectivity counterexample belongs in the story
A PARP design study reports a selective compound and another with the opposite preference. Both deserve a place in the optimization record.
When solubility prediction needs a better measurement
An organic-solvent study makes the quality of the test data part of the model-performance story.
An energy-storage number needs its electric field
A ceramic inverse-design study links composition, polarization patterns and four rounds of synthesis. The fair comparison keeps the test conditions attached.
A good spectral match can still name the wrong molecule
NMR-Solver links structural proposals to predicted chemical shifts. A reported failure exposes why that link needs an independent check.
An antibody must reach its target
A simulation-plus-learning study asks which combinations of exposure and target biology sustain engagement.
When an AI prescription becomes a real intervention
A prospective transplant study shows conditional prescribing in practice. Feasibility and treatment benefit remain different questions.
Who chooses the compartments?
Automated pharmacokinetic modelling makes the search more explicit. Its useful output is a defensible model of exposure.
The drug-interaction predictions left outside the confident set
T-DDI reports stronger performance on a selected subset. The remaining predictions are part of the pharmacology story.
Listening to liver cells before calling a drug safe
ToxPredictor connects human-hepatocyte expression patterns with exposure. Its validation denominator belongs beside the headline.
A drug response needs a clock
XPert treats concentration and time as part of the prediction, bringing cellular response closer to a pharmacological question.
Teaching a language model to carry potency across scaffolds
A new Sanofi study turns examples from established chemical series into compounds tested at the bench. Its value emerges in the relationship between prompting, selection and measured inhibition.