Confidence-Weighted Elastic Gaussian Networks To Predict Protein Flexibility

Rapid protein flexibility prediction is useful for evaluating novel folds generated by deep learning models. Gaussian Network Models (GNMs) provide a physics-inspired framework for this task but assume uniform spring constants, ignoring the per-residue confidence information available from modern structure predictors. We introduce a zero-parameter modification that weights the GNM spring constants with AlphaFold2 predicted Local Distance Difference Test (pLDDT) scores and inverse squared distance: $\gamma_{ij} = p_i p_j / d_{ij}^2$. On the ATLAS molecular dynamics benchmark (1932 proteins), the mean Pearson correlation increases from $r = 0.765$ to $r = 0.841$. Partial correlation analysis confirms that pLDDT contributes information beyond local geometry. A lightweight Graph Neural Network trained to correct analytical residuals reaches $r = 0.871$, approaching the inter-replica ceiling of $r = 0.88$. These results suggest that pLDDT encodes mechanical information that is not captured by contact geometry alone.