What Cycle Closure Can and Cannot See in a Relative Binding Free Energy Network

Résumé fourni par la source

Cycle closure is the standard consistency check on a perturbation network, and what such a check can see is fixed by the gradient–cycle decomposition of edge flows (HodgeRank): the cycle part of a network’s error is auditable and biases nothing, while the gradient part biases the affinities and is invisible to any statistic computed from the network’s own values, at any magnitude. Both parts are measured here on 1145 real binding edges. The auditable subspace spans a third of the error space and, in the variance-weighted metric the closure statistic uses, carries under one per cent of the error against experiment. Experimental and annotation error is per-ligand, so it falls in the invisible part by construction and can only have made that share smaller: the measurement is a floor, and it survives correction: in the metric the closure statistic uses on every system measured, at one cell as a tie, and in an isotropic reading on every system of the wider grounded set, failing only on the two weakest of the four the test itself selected. The test built on it behaves as the bound predicts, and its flag set is not stable: the per-edge residuals and the ranking of systems reproduce across independent repeats, and the list of flagged systems does not. Fitting closure against the precisions the estimator already reports gives that test its stated null; it ties the standard fixed 1.0 kcal/mol cutoff, and on every replicate the set of systems it flags lies inside the set that cutoff flags. It flags nothing once the reported errors are widened by half, enough to cover the pose and preparation variance the protocol repeats hold fixed, and it names no removal that recovers accuracy against experiment. Resolving the gradient part needs measurements from outside the network, at worst one per unresolved dimension.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.