Flexibility or uncertainty? A critical assessment of AlphaFold 2 pLDDT

Release of AlphaFold 2 and subsequent development of AlphaFold 3 had a profound impact on protein structure prediction, providing near-experimental accuracy. However, the utility of AF2's confidence index (pLDDT) as indicators of protein flexibility remains underexplored and debated. In this large-scale study, we evaluate AF2's pLDDT as a predictor of protein flexibility by comparing it with flexibility metrics derived from molecular dynamics (MD) simulations from the ATLAS dataset, NMR ensembles, and experimental B-factors. We also assess the efficiency of ESMFold pLDDT and AlphaFold 3 in this context. Our findings reveal that AF2 pLDDT reasonably correlates with MD and NMR-derived flexibility metrics, but fails to capture flexibility in the presence of interacting partners, and therefore need to be cautiously interpreted. Furthermore, AF2 pLDDT appears more relevant than B-factor values for evaluation of protein flexibility. While AF3 shows slight improvements in capturing protein dynamics, MD simulations remain superior for comprehensive flexibility assessment.