Metaheuristic Optimization of Temporal Surrogates for Early Prediction of Protein Conformational Stability from Short Molecular-Dynamics Trajectories
Year: 2026
Type: article
Abstract: Metaheuristic optimization can strengthen recurrent surrogates, but its value must be judged against controls and hyperparameter optimizers under leakage-free evaluation. We predict five equilibrium protein observables from 20 ns prefixes of 100 ns molecular-dynamics trajectories using a gated recurrent unit (GRU) configured by an Improved Grey Wolf Optimizer (Improved-GWO). The benchmark comprises 20 ATLAS proteins (60 trajectories) under nested leave-one-protein-out cross-validation. Relative to the original fixed GRU, Improved-GWO had lower absolute error for several observables; Holm-adjusted paired tests remained below 0.05 for SASA, helix and strand. Among six models, last-value persistence achieved the lowest nominal MAE for all five endpoints, and no persistence–Improved-GWO comparison survived Holm correction. A fixed 512-unit GRU improved nominally over the original small GRU, while paired tests favored Improved-GWO for helix and strand. In matched-budget HPO, Bayesian Optimization attained lower best inner-validation fitness than Improved-GWO, which in turn was lower than Random Search. Configurations from Random Search and Bayesian Optimization were not evaluated on outer tests. Ridge remained competitive. Thus, the benchmark does not establish a consistent predictive advantage from recurrent complexity over temporal or regularized linear baselines. The p53/NAMD assessment remains preliminary zero-shot evidence, and trajectory-length reduction is not a measured wall-clock speedup or validated stopping policy.
Source: Mathematics
DOI: 10.3390/math14203679
Institution: Universidad de Valladolid
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7221567829