Why should a predicted structure not be used alone to select mutations?

Why

AlphaFold can give a locally confident fold while offering limited information about alternative conformations, ligand-bound states or mutation-driven changes in free energy. A high pLDDT residue may still be part of a flexible catalytic region whose dynamics the static model cannot reveal. Use predicted structures to generate testable hypotheses, inspect local confidence and predicted alignment error, and compare known functional residues or experimental structures when available.

Consider conserved sequence patterns and structural contacts, but avoid assuming that an apparently strengthened hydrogen bond improves the protein. For example, mutating a loop to increase predicted packing can obstruct substrate access or reduce expression. Build a small panel of plausible alternatives, confirm identities and compare stability, solubility and function experimentally.

When confidence is low, increase empirical exploration rather than inventing precise mechanistic predictions from a single image. Predictions are a guide for prioritization, not experimental validation.