Figure 4.

Histograms displaying bond-length distributions for the 24 bond classes with at least 100 observations that have the highest multimodality indices (a modified version of Sarle’s coefficient). The title of each histogram identifies the atom pair corresponding to the bond class. All histograms are shown on the same scale (1.2–1.55 Å). (a) corresponds to AceDRG-derived data generated in June 2014 and (b) to data generated in January 2015. In the earlier table, it is evident that there were a number of classes that clearly exhibited multimodal behaviour and comprised bond-length observations spanning a wide range of values. This indicates the original atom typing to be insufficient to describe the local chemical environments in a way suitable for purpose. In the latter table, even the most extremely multimodal classes span a comparatively shorter range of values, inevitably resulting in smaller derived standard deviations, and thus more accurate restraints. The class table from June 2014 comprised 268 882 bond classes, of which 2996 had at least 100 observations. In contrast, the class table from January 2015 comprised 169 362 bond classes, of which 1222 had at least 100 observations. This illustrates how the selection criteria have become more stringent from one table generation to the next. This figure was produced by R (R Core Team, 2013 ▸).