Density-guided AlphaFold reveals unmodeled alternative turn conformations in protein structures

Protein structures derived from X-ray crystallography often underrepresent local conformational heterogeneity despite clear support from experimental electron density. Here, we introduce an electron density--guided AlphaFold approach for generating local multiconformer ensembles of $\beta$-turn regions directly from crystallographic density maps. The method combines density-guided conformational sampling with dihedral-angle-based classification to filter ensembles and identify distinct clusters of backbone conformations consistent with the experimental density. Across multiple case studies, ED-guided modeling recovers known alternative states, resolves cases of ambiguously deposited altlocs, and uncovers additional conformations absent from deposited PDB models. These results suggest that experimental density-guided structure generation can enable large-scale studies of local conformational variability across the PDB, providing a route to systematically characterize heterogeneous structural states that are only partially captured in the PDB.