Validating AlphaFold2 structure predictions with cross-linking mass spectrometry of endogenous ciliary proteins

Advances in protein structure prediction have taken the biology community by storm with their promise of readily available, high-quality, atomic-level structural information. Because a major goal in structural biology is to understand proteins in their native biological contexts, we asked if AlphaFold2 predictions match native protein conformations. To answer this question, we chemically cross-linked proteins within intact Tetrahymena thermophila cilia and fractions enriched for endogenous protein complexes and identified cross-linked amino acids by mass spectrometry. Across the 100 best-sampled proteins, we measured 1,228 intramolecular cross-links, which serve as a benchmark of distance restraints to be obeyed by endogenous proteins. The corresponding AlphaFold2 structure predictions were highly concordant, positioning 86.0% of cross-linked lysine residues within 30 Å of each other (Cɑ-to-Cɑ), consistent with the known size of the cross-linker, and 92.2% within 40 Å of each other. A large fraction (43%) of the proteins showed no violations at all. Meanwhile, most inconsistencies occurred in low-confidence regions or not within confidently predicted rigid domains. For basal body EF-hand protein BBC118, the cross-links could be combined with the predicted structure to find domain packing satisfying both sets of data. Overall, AlphaFold2 performs well at predicting biologically relevant protein structures, with high confidence values corresponding to more correct native structures. However, we observe cases involving incorrect relative orientations of rigid domains, signifying the importance of combining structure prediction with experimental information to reveal biologically relevant structural conformations.