One of my proudest contributions: https://www.ncbi.nlm.nih.gov/pubmed/24265211
Protein design and fold prediction was being held back by using fixed (equilibrium) bond angles in proteins, and if you have enough CPU time, you can optimize those values (they often end up centering around non-equilibrium angles for the particular atom types) to get better predictions.
Do you happen to know if this is related to what was done with AlphaFold? It sounded like they moved from fixed values to a probabilistic model but I'm not sure if bond angles was the property they were measuring.
Great question. I don't know.
To put numbers on this, a typical approximation for bond energy as a function of stretch would be E = 500x^2, where x is the stretch in Angstroms and E is the energy change in units of thermal noise at room temperature (about 0.6 kcal/mol). A bond length deviation of 0.01 Angstroms is already negligible, because its associated stretch energy is far below thermal fluctuation levels. Parts per billion are totally invisible.
Bonds are flexible and can hold mechanical stress, so there is only a "preferred" length, not a guaranteed length.