AI chemistry works when something outside the model checks it

The same lesson in four preprints

Four ChemRxiv preprints this cycle test AI methods on chemistry problems that look very different: designing graphene membranes for desalination, screening libraries of tens of millions of compounds, swapping the core of a drug molecule, and predicting molecular properties. They share one lesson. In each case the AI method works because something outside the model checks it: a predictor that scores each design, an oracle that labels each pick, crystal contacts that a replacement must preserve, or the chemistry of non-covalent interactions built into the model. All four are preprints, so none of this is peer-reviewed yet. But the pattern is consistent enough to matter for anyone building these tools: the model is the junior partner, and the feedback loop does the work. m:…6540 m:…1347 m:…4276 m:…9435

Membrane design only works with feedback

The desalination preprint is the clearest test of the lesson. LLM agents were asked to design graphene nanopores for water desalination, and the headline numbers look strong: Claude Opus 5.5 achieved 122.3 ns-1 flux at 93.2% rejection, while GPT-6.1 Sol achieved 131.7 ns-1 flux at 90.2% rejection. The preprint's own conclusion undercuts any reading of that as raw model ability. Predictor feedback is the primary determinant of LLM agent performance in nanopore design, as models without feedback failed to maintain ion-rejection floors. A greedy search restricted by simple manufacturability rules matched or exceeded LLM agent performance, according to the preprint. The implication is blunt: for this design task, a simple search plus a good predictor beats a frontier model left to reason on its own. m:…6540

Screening millions by testing thousands

The other two preprints apply the same logic to drug discovery. LearnM8 1.0 is a modular active learning framework that plugs in any oracle, from docking to laboratory measurements, and it ships with 19 learners, 7 acquisition strategies and 39 featurizers. In retrospective benchmarks it recovered 98% of the top 0.1% of compounds in AmpC libraries of up to 96 million compounds after evaluating only 10% of the library, and 91% of the top 0.1% in a 116-million-compound D4 library after evaluating 10%. Random selection recovered 9.4% of the top 0.1% in a 1-million-compound AmpC library, which sets the baseline the framework has to beat. The scaffold hopping preprint works from a different kind of ground truth: a docking-free contact-conservation benchmark built from 2,886 PDB ligand pairs. Its SV-Hop method, which pairs a 60,069-rule core-replacement lexicon with a learned contact ranker, achieved a 62.1% success rate in scaffold-held-out references, against 50.3% for property-based methods and 44.3% for 2D similarity. True hops recovered 80.6% of reference contacts across 40 crystal pairs, which is the standard the method is judged against. A smaller 388-rule PDB-only lexicon reached 81.9% success on target-held-out splits, outperforming the full 60,069-rule lexicon at 65.3%, a reminder that more rules are not always better. m:…1347 m:…4276

Chemistry built into the model

The fourth preprint builds the grounding into the model itself. ChemPrior-GAT improves molecular property prediction by explicitly modelling non-covalent interactions and pharmacophoric roles rather than relying on distance-based cutoffs. It achieves the best ROC-AUC on three MoleculeNet classification benchmarks, and its rule-based non-covalent graph is approximately 20 times sparser than a 5-angstrom cutoff graph, which keeps the model focused on interactions that chemistry says matter. The preprint also reports prospective experimental validation: on 16 compounds, the model achieved a mean absolute error of 0.50 log units for logD7.4. The pharmacophore-supervision module is transferable to other graph neural network architectures, such as AttentiveFP, which matters because it means the idea can travel without rebuilding the whole model. m:…9435

What to watch

The practical takeaway for builders is about where to spend effort. The desalination result says to budget for the predictor and the feedback loop before paying for a bigger model, since a greedy search with good feedback matched the agents. The LearnM8 and SV-Hop results say that the quality of the oracle and the benchmark decides whether the method is real: active learning is only as good as the labels it buys, and scaffold hopping is only as good as the contacts it preserves. The ChemPrior result says that encoding known chemistry into the model can beat adding parameters. What is still missing is scale of validation: the prospective test for ChemPrior covered 16 compounds, the desalination designs are simulated rather than built, and the screening gains are retrospective. The next signal to look for is prospective, experimental confirmation at larger scale. m:…6540 m:…1347 m:…9435