Predicting Molecular Dynamics Contact Persistence from Static Structures for Weighted Gaussian Network Models
TL;DR
Persistence is a useful and auditable edge target, but the present mapping recovers only a modest, protein-dependent improvement rather than a new state of the art.
Abstract
Uniform Gaussian network models (GNMs) are inexpensive and interpretable, yet they assign the same coupling to contacts that behave quite differently in molecular dynamics (MD). This work examined a narrower question: can the fraction of MD frames in which a contact survives be inferred from one static structure and then used as a GNM weight? Contact-persistence and Cα root-mean-square fluctuation (RMSF) labels were reconstructed from three ATLAS replicas for 200 protein chains. Evaluation used an Evolutionary Classification of Protein Domains (ECOD) component-disjoint split of 132/38/30 proteins. On the 30-protein test set, the validation-locked histogram gradient boosting (HGB) model reached a protein-macro mean absolute error of 0.0856 and a Spearman correlation of 0.8878. Most of that predictability came from geometry: the margin over distance-isotonic regression was statistically supported but small. With predicted persistence mapped to nonnegative E8 couplings, RMSF Spearman correlation rose from 0.7170 for uniform GNM to 0.7229. The component-level gain was 0.00536 (95% bootstrap confidence interval, 0.00126-0.00962; Holm-adjusted sign-flip p = 0.0173 ). A continuous-persistence oracle reached 0.7657. Thus, persistence is a useful and auditable edge target, but the present mapping recovers only a modest, protein-dependent improvement rather than a new state of the art.