Recent
breakthroughs in neural network-based structure prediction
methods, such as AlphaFold2 and RoseTTAFold, have dramatically improved
the quality of computational protein structure prediction. These models
also provide statistical confidence scores that can estimate uncertainties
in the predicted structures, but it remains unclear to what extent
these scores are related to the intrinsic conformational dynamics
of proteins. Here, we compare AlphaFold2 prediction scores with explicit
large-scale molecular dynamics simulations of 28 one- and two-domain
proteins with varying degrees of flexibility. We demonstrate a strong
correlation between the statistical prediction scores and the explicit
motion derived from extensive atomistic molecular dynamics simulations
and further derive an elastic network model based on the statistical
scores of AlphFold2 (AF-ENM), which we benchmark in combination with
coarse-grained molecular dynamics simulations. We show that our AF-ENM
method reproduces the global protein dynamics with improved accuracy,
providing a powerful way to derive effective molecular dynamics using
neural network-based structure prediction models.