Daily digest — 2026-11-14

No new wet-lab binding data here: scoring against existing fitness datasets and in silico predictors invites circularity, especially if related models score the designs. Still, one latent space where off-target avoidance is just a direction is appealing. Next step: SPR or ITC on a handful of designs, including the off-target-aware ones, to see whether the selectivity gap survives.

Independent, additive pixels are a convenient fiction: real SAR is full of non-additivity, conformational coupling and activity cliffs, so the size estimate is only as good as that assumption. And DEL read counts are noisy proxies for affinity. Before trusting any library-size formula, I'd compare it against simply synthesizing a focused hit-centered set off-DNA and measuring it.

Would we trust a generator that is rewarded by the very classifier we'd use to judge it? Reward hacking is the usual failure here: the policy finds sequences the language model loves but membranes don't. Before this touches a real project, I'd want wet-lab uptake and toxicity data on a handful of designs, plus a check for endosomal trapping, which classifiers on CPP databases rarely capture.

Monday: pull CROWN and check its similarity annotations against our own targets before retraining anything. A CASF-2016 win is weak evidence, since that benchmark is saturated and overlaps heavily with PDB training data. The parity on docking and screening is the more honest signal. Retrospective only; what matters is whether repaired structures help on our series.

The family-level split is the only one of the three schemes that says anything about novel targets; random splits leak near-identical complexes across folds. Even so, PPI ΔΔG sets are small, redundant and skewed toward destabilizing alanine mutations, so a gain over 13 baselines may not survive contact with FEP or empirical tools on a real affinity-maturation campaign. Prospective designs are the missing test.

If you need to flag off-target pockets for a selectivity screen, an alignment-free matcher is the right tool, and the distance-as-function idea sidesteps the arbitrary cutoff problem that plagues fpocket-style descriptors. But three retrospective benchmarks say little about apo or induced-fit pockets, and I'd want a head-to-head against SiteEngine and graph-based comparison on my own target family before trusting it.

Most of this will be familiar to anyone who has watched a model flip on a re-enumerated SMILES, but the perturbation-test idea is the usable part. On our own project I'd run it before trusting any benchmark gain: randomize SMILES, swap tautomer or protonation states, then see whether error tracks the representation or the model. Without prospective data, though, a Perspective can't tell us which fixes matter.

Before adopting any off-the-shelf embedding, I'd open the tokenizer and the fine-tuning scripts, not the model card. If the downstream benchmarks were run on non-isomeric SMILES, they can't show whether the model resolves chirality, so reported gains say nothing about stereo-driven SAR. I'd add enantiomer and diastereomer pair tests to our own evaluation first.