Daily digest — 2026-11-10

If the training data is MD, matching MD distributions is a low bar. The harder question is whether Cadense recovers CDR-H3 states that long simulations rarely visit, and how it handles antibodies far from its training set. With no experimental check reported, try it first on a sequence-held-out set of paired apo/holo structures before trusting ensemble docking.

Generating frames autoregressively is sensible, since independent-frame generators discard kinetics. But MD surrogates should be judged on slow-state populations and transition rates against long reference runs, not per-frame fidelity. For protein-ligand systems, check whether unbinding or pose switching appears at all, and whether long rollouts drift. Without code or held-out systems, this is promising rather than usable.

A single importin-α/NLS system with only moderate ITC agreement is one data point and does not establish the protocol. The entropy-driven reading depends on a water and ion decomposition that shifts with force field and ion parameters. Compare the computed enthalpy directly with calorimetry, and benchmark against standard double decoupling or MM-GBSA on the same system.

Ligand parameters are usually the weak link in capsid simulations, so twelve inhibitors plus IP6 in one CHARMM-compatible set is a practical gain. What remains open is whether the lenacapavir free-energy estimates hold for compounds left out of parametrization, since agreement with known data alone is a fit. Test how well the coarse-grained mapping transfers to your own analogs.

That AMOEBA samples reactive geometries and AMBER does not is intriguing, yet it proves little on its own: a polarizable model can over-stabilize near-transition-state conformers as easily as it corrects a fixed-charge bias. What is missing is a comparison with experiment, such as a QM/MM barrier or kcat shifts in known mutants. For now it argues for budgeting polarizable runs on electrostatics-driven targets, not for a verdict.

Permutation controls show only that the models beat noise, not that the split blocks analog leakage, which matters because ChEMBL series are dense with near neighbours. Unless splits were scaffold- or time-based, expect optimistic numbers. Thirteen benzylidene compounds in MCF-7 cells is a single-chemotype, cell-level check that cannot validate target-specific predictions. Per-target results against a plain Morgan-fingerprint random forest baseline would help.

A scaffold split still leaves close analogs across the boundary, so a tuned nearest-neighbor matching the neural net says as much about the benchmark as the models. The blend's gain is small, so check effect size per target, not just significance. A prospective test would be more telling: how do both rank targets for a novel chemotype with sparse ChEMBL coverage?