How should AlphaFold predictions be interpreted and what are its limitations?

AlphaFold2 interpretation requires understanding its outputs and limitations: 1) Confidence metrics - pLDDT (per-residue confidence, >90 high, 70-90 good, <50 likely disordered); PAE (predicted aligned error) indicates domain relationships and confidence in relative positions. 2) Strengths - excellent for globular domains with homologs; accurately predicts backbone and most side chains. 3) Limitations - predicts single static structure, not conformational ensembles or dynamics; may not capture effects of ligands, post-translational modifications, or partner proteins; struggles with intrinsically disordered regions, membrane proteins in lipid context, and large conformational changes; novel folds without homologs have lower accuracy. 4) Multi-domain proteins - domains may be correctly folded but relative orientations uncertain (check PAE). 5) Complexes - AlphaFold-Multimer addresses this but accuracy varies. Best practices: validate with experimental data, use molecular dynamics for dynamics, be cautious with low-confidence regions.