1In the present case, however, ADPs are kept fixed and blocks of parameters are not refined in turn.

2QM/MM stands for quantum mechanics/molecular mechanics and MO/MO stands for molecular orbital/molecular orbital, where different methods/basis sets can be combined. Svensson et al. (1996BB39) introduced ONIOM as an abbreviation for `our own n-layered integrated molecular orbital and molecular mechanics'. The connotation to the shell-like structure of an onion was most likely intended.

3ADPs refined from neutron diffraction are often reported in the literature, while intensities from neutron diffraction are usually not available.

4Supplementary data for this paper are available from the IUCr electronic archives (Reference: TN5019 ). Services for accessing these data are described at the back of the journal.

5Even when neutron data are used, setting X-H bond distances can improve convergence behaviour.

6The Gaussian03/09 option opt = CalcFC improves the convergence behaviour and is highly recommended. Usage of tight or very tight convergence criteria is recommended, but does not change ADPs by much.

7Using all six elements of the Uij's gave nearly identical results.

8The scale factor could of course also be obtained without any experimental input, since an empirical procedure where scale factors are determined by systematic variation is also feasible. Nevertheless, since these ADPs are used in refinements on experimental data, they are not independent of experiment in a strict sense, even more so since phonon modes are not predicted in cluster calculations.

9The dipole moment of the whole cluster seems to correlate with the quality of ADPs. For both L-cysteine and L-threonine the dipole moment is minimal for the 15-molecule cluster, for L-alanine it is the second lowest. We assume that, not only for zwitterions, a low cluster dipole moment is characteristic and indicative of a balanced environment.

10For MP2 calculations the Gaussian keyword density = current needs to be used.

11Ochterski (1999BB32) recommends choosing a fine grid for DFT calculations, as is the default in Gaussian09. An ultra-fine grid was tested [option Int(Grid = UltraFineGrid)] for L-alanine and L-cysteine. Differences in ADPs and scale factor are detectable but small.

12PM3, AM2 and HF/STO-3G did not allow us to consistently reach convergence in Gaussian03 and were therefore discarded. We recommend the use of Gaussian09 for ADP calculation.

13For L-threonine, MP2 calculations already required temporary disk space of [\simeq] 1 TB and RAM memory of 24 GB. Computation time took weeks on a current Xeon 8 cpu server. We therefore decided that the computational requirements for MP2 MO/MO calculations are currently too high for a method that is supposed to be computationally efficient.