Daily digest — 2026-11-15

Two validated binders out of 1,600+ predictions is a hit rate near 0.1%, and that is the honest number for a target with little chemical matter. Note the winner leaned on patent inhibitors, so the useful lesson is that known-ligand data still beats pure structure-based scoring. Next step for a team: check whether orthogonal assays, not one readout, confirmed the two hits.

Better docking scores on CrossDocked2020 are the weakest possible win: docking is the very signal the memory module learns from, so the loop can simply overfit the scoring function, and the benchmark's pocket overlap with training data muddies generalization. Ask for comparison against a pocket-conditioned diffusion baseline at matched compute, plus MM-GBSA or FEP rescoring on a few targets. Without that, it is a docking optimizer with an LLM wrapper.

Inter-study disagreement on identical compounds is the number to watch: differing measurement methods, pH, ionic strength and counterions can swamp any model gain. Splitting by article is the right leakage control, but no prospective synthesis-and-measure test is reported. A team could use the series-calibration trick directly: measure two or three analogues, then rank the rest.

Synergy models rarely fail on random splits; they fail when the drug or cell line is new, so the unseen-drug tests are the ones to read first. DrugComb-style labels are noisy and batch-dependent, so check for leakage across overlapping datasets and compare against a tuned gradient-boosting baseline. Without wet-lab confirmation of top-ranked pairs, treat this as a prioritization aid, not evidence.

"Comparable to published models" is the number to interrogate: if a fingerprint model on the same cell-assay data does equally well, the simulation cost buys only interpretability. That per-group energy attribution is the real appeal, but it needs a scaffold-split or time-split test and a held-out series from your own project before you trust it. Transporter-confounded cell-assay labels will cap any passive-diffusion physics.

Contact-area weighting is the idea worth testing: it captures buried surface and packing that plain distance cutoffs miss, and Voronoi cells are cheap to compute. But binding-site benchmarks are prone to leakage through homologous or shared-ligand proteins, and the gap to P2Rank or recent GNN predictors is often small. Check sequence-identity splits and apo or AlphaFold structures before trusting the headline.

One in vitro candidate cannot validate a design loop whose fitness function is the same model being reported as accurate. A GA with simulated annealing will exploit any blind spot in the scorer, so higher predicted affinity over the parent is expected by construction. Ask for SPR or BLI on a dozen designs, spanning high and mid scores, and a comparison against ESM-style likelihood and structure-based baselines.

Peptide ADME data is small, noisy and scattered across assay protocols, so a single model spanning Caco-2, PAMPA, half-life and clearance deserves a hard look at split design: random splits on near-identical sequences inflate scores. Ask whether the test sets were clustered by sequence or scaffold, and whether it beats a simple baseline on your own cyclic or modified series before trusting any number.