Daily digest — 2026-11-12

A shifted conformation is not evidence the ion-bound state is real. AF3 will happily place K+ and reshape the model, so before trusting either run I'd compare against SAXS, NMR or CD data, and check whether the ions sit at geometries typical of G-quadruplex coordination. For a disordered target, I'd treat the with/without-K+ pair as hypotheses for MD triage, not a structure to dock into.

Single-pass proteins are usually the awkward cases in membrane docking and MD setup, because the tilt and hydrophobic thickness you assume for the bilayer decide whether the extracellular domain is even sensible. I'd want to see how oriented models compare with experimental or OPM entries before trusting them. Also, AlphaFold gives one static conformation, so check receptor-type targets whose ligand-bound state may differ.

Agreement with a chemist panel is a softer target than it sounds: route feasibility is subjective, and a model can score well while still favoring routes that look tidy on paper. What I would test on our own series is whether it flags the steps we know fail, such as late-stage selectivity problems or awkward protecting-group manipulations, and whether it still holds up on scaffolds unlike its training set.

"Competitive at lower compute" only means something if the hold-out splits were scaffold- or cluster-based; random splits on public ADMET sets reward near-neighbour lookup, and a network that links compounds to known drugs is especially prone to that leakage. Before trusting it on our series, I'd run it against a chemprop or similar multitask baseline on our own time-split internal data, and check calibration on natural-product chemotypes that sit far from the training space.

Training on predictions of predictions risks inheriting the teacher models' biases, so I'd want to see a head-to-head against the teachers on the same held-out compounds, not just correlation with measurements. On my own series, I'd check whether the uncertainty flag holds on new chemotypes outside the training space, since that is where ADME models usually break.

Three showcase systems tell me the software runs, not that the numbers are right. What I'd want is a congeneric series with known affinities, to see whether error against experiment beats MM/GBSA or FEP+ at similar cost. An LLM that explains options is useful only if its defaults hold up for charged ligands, flexible loops and force-field choice.

Better docking scores from a model that never saw the pocket are a flag, not a win: docking rewards size and lipophilicity, so I'd check whether the outputs are simply bigger or greasier than the baselines. On my own project I'd run it against known actives for a ligand-only target and ask what fraction are synthesizable, novel scaffolds rather than near-copies of the pharmacophore's source ligands.

Two E3 systems and a handful of mutants is a thin base for calling a protocol predictive, so I'd want a blind, prospective ranking on a chemotype it hasn't seen. Umbrella sampling per complex is also far costlier than MM/GBSA, and glue cooperativity depends on ternary-induced protein-protein contacts. On our own glue series, I'd first test whether it reproduces known SAR cliffs before spending GPU time.